A Study on the Performance of Actively and Passively Managed Artificial Intelligence Exchange Traded Funds

Abstract

This study employs a sample of 25 active and 22 passive AI ETFs to examine several issues surrounding their performance, risk, pricing efficiency, and persistence in pricing discrepancies and their impact on ETFs’ performance combined with the respective impact of intraday volatility. The relationship between AI ETFs’ performance and market factors concerning size, value, profitability, investment and momentum is evaluated too. The results indicate that the passive AI ETFs have outperformed active ones over their entire trade history, without, however, shouldering their investors with materially higher volatility. Moreover, both AI ETF groups trade at a persistent premium to their NAV. The concurrent premium positively affects return, while the one-period lagged premium is negatively related to return. In addition, a negative relationship between return and concurrent intraday volatility and a positive (but less strong) relationship between return and one-period lagged intraday volatility are found. Moreover, the majority of AI ETFs do not generate significant alphas. Finally, market factors effectively explain the performance of AI ETFs.

1. Introduction

Artificial intelligence (AI) refers to the capability of computers or computer-controlled robots to perform tasks that are typically related to human intelligence, including comprehension, learning, reasoning, problem solving, decision making, creativity and autonomy. As noted by Moor (2006), the field of AI research was founded at a workshop in the Dartmouth College in 1956, where John McCarthy coined the term “artificial intelligence”. Turing (1950) raised the question of whether “machines can think”. In the early 2000s, machine learning was being used in a wide range of applications in academia and industry, facilitated by the availability of powerful computer systems and the significant evolution in data collection and storage, as well as the application of solid mathematical models.1 However, according to Zysman and Nitzberg (2020), investment in AI boomed in the 2020s. This boom led to the rapid scaling and public releases of large language models like ChatGPT.

Apart from being in the core of the fourth industrial revolution, AI offers investors very promising investment opportunities, such as AI Exchange Traded Funds (AI ETFs).2 These ETFs focus on firms that, in some way, participate in the artificial intelligence sector, providing investors with diversified exposure to this fast-growing industry at a relatively low cost compared to individual stock investments. These ETFs invest in a wide range of firms that focus on robotics, self-driving vehicles, large language models, generation or improving the manufacturing processes, while tapping into the prospects of a technology that has just begun to scale. In comparison to AI mutual funds and index funds, AI ETFs are more flexible due to their continuous trading, also exceling in liquidity and cost terms.

From another perspective, the term AI ETFs often refers to ETFs employing artificial intelligence algorithms to assist the selection process and the management of the underlying portfolio. However, the assistance in the portfolio selection process by AI tools does not entail that AI securities are finally selected. In any case, ETFs that are built around AI stocks are not all the same. AI ETFs differ from each other in several ways, concerning fund size, price, risk, underlying value, and dividend policies. In addition, they may invest in different stocks or indices. One key differentiating element concerns the management approach of AI ETFs, that is, whether these ETFs adopt active or passive management strategies.3 In fact, more AI ETFs pursue active management than those adopting a passive management approach.4

An important factor to consider when choosing to trade with AI ETFs is the risk of these investment products. Similar to general investment risks, investing in an AI ETF could result in money loses and, obviously, past performance cannot guarantee future returns. In addition, similar to any emerging technology, there is a risk that some of the respective companies may go out of business.5 AI business is a highly competitive and volatile industry, with frequent changes in the synthesis of the sector’s key players. However, as already noted, due to the diversification techniques adopted by AI ETFs, these funds are considered to be less risky than investing in individual AI stocks.

In this study, we examine a series of aspects concerning trading with AI ETFs. More specifically, we gather a sample of 25 active and 22 passive AI ETFs with trade data of at least one full year up to 31 December 2025, and focus on their performance, risk, and price efficiency, that is the discrepancies between the trade prices and Net Asset Values (NAVs) of ETFs, the persistence in these discrepancies and the possible impact on returns, along with the respective impact of intraday volatility. The relationship between AI ETFs’ performance and key market factors concerning size, value, profitability and investment patterns, as prescribed by Fama and French (2015), and the momentum factor suggested by Carhart (1997), is examined too.

The results show that, on average, passive AI ETFs have performed better than active ETFs at the cumulative level over their entire trade history, both in trade price and NAV terms. On the other hand, the average risk and intraday volatility records of the two groups are quite close to each other. Moreover, active and passive AI ETFs are found to trade at an average daily premium to their NAV of 4 and 7 basis points (bps), respectively, while both groups present a premium during 60% of their trade history. Via regression analysis, the premium is found to be quite persistent. At least, persistence can go back up to five trade days. On the question of whether the premium can affect the return of AI ETFs, the empirical analysis shows that the concurrent premium is more important in explaining returns (in a positive way) in the case of passive ETFs than in active ETFs, while the one-lagged premium is negatively related to the returns of both groups. The analysis also reveals a negative relationship between return and concurrent intraday volatility, both for active and passive AI ETFs, and a positive (but less strong) relationship between return and one-lagged intraday volatility. Finally, on the relationship between AI ETFs’ performance and market factors, the regression analysis reveals that only a limited number of AI ETFs produce significantly positive alphas over the Russell 3000 Index. Furthermore, the performance of AI ETFs is related to Fama and French’s (2015) size factor and Carhart’s (1997) momentum factor in a positive way (on average terms), while performance is negatively related to the Fama and French (2015) value and robustness factors. The impact of the conservativeness factor is less significant.

This study has been motivated by the general growing interest in AI applications and the prosperous potential of the entire AI industry for the years to come, as well as the increasing interest in AI ETFs in particular. This study is not the first to focus on AI ETFs. For instance, Wu and Chen (2022) investigate the impact of AI ETFs on their underlying stocks. Bonaparte (2024) assesses the financial opportunities arising from the new AI innovation and proposes a valuation model for the respective ETFs and stocks. Chen and Ren (2022) evaluate the performance of AI-powered mutual funds, a cousin to AI ETFs. Poutachidou and Koulis (2025) study the performance of ETFs investing in firms involved in AI technologies by considering the impact of active management. Malhotra (2025) evaluates the performance of AI (and Blockchain) ETFs, trying to identify whether these funds offer opportunities for generating positive alphas.

To the best of our knowledge, no study examines the pricing efficiency of AI ETFs by focusing on the differences between the trade prices and NAVs of these funds. Nor does any study assess the impact of these price discrepancies on the return of AI ETFs. The impact on the return of AI ETFs of their intraday volatility has also been neglected in literature. In addition, even though the merits of active management for the performance of AI ETFs have been assessed in the literature, the impact of active management on their pricing efficiency has not been evaluated yet. Our study seeks to fill these gaps. We deem our results to be quite important given the statistically significant relationships revealed between AI ETFs’ returns, premium and intraday volatility. Investors could benefit from our findings by cashing in on these relationships.

The rest of the paper is organized as follows: Section 2 establishes the theoretical framework relating to the results of our study. Section 3 provides an analysis of the key findings of literature concerning AI ETFs. Section 4 breaks down the research objectives of the current study. Section 5 describes the research methodology and the sample of our study. Section 6 discusses empirical findings. Summary and conclusions are offered in Section 7.

2. Theoretical Framework

Although being purely empirical, this study contributes to the growing literature on thematic investing by explicitly positioning the performance of active and passive AI ETFs within four core frameworks of financial economics: (i) the Efficient Market Hypothesis, (ii) ETF arbitrage theory, (iii) behavioral finance, and (iv) modern asset pricing.

From the perspective of the Efficient Market Hypothesis (EMH), our analysis evaluates whether return dynamics in AI ETFs are consistent with semi-strong informational efficiency. Under this framework, abnormal returns, that is statistically significant positive alphas, should not persist once publicly available information is incorporated into prices. Our empirical findings show that the majority of AI ETFs do not generate statistically significant positive alphas, suggesting that excess returns associated with AI investing are largely competed away. This evidence is consistent with semi-strong market efficiency, even within a rapidly evolving and innovation-driven sector such as artificial intelligence. At the same time, the presence of short-term deviations, such as ETF premiums, indicates that efficiency may be conditional and subject to frictions rather than absolute.

Second, we relate our findings to ETF arbitrage theory, which predicts that the creation and redemption mechanism implemented by authorized participants should enhance price alignment between AI ETF trade prices and their underlying Net Asset Values. In a frictionless market, any deviation from NAV would trigger arbitrage transactions to quickly eliminate mispricing. However, our evidence of persistent premiums in both active and passive AI ETFs suggests that arbitrage is not fully effective in this context. We interpret this through the lens of limits to arbitrage, including transaction costs, liquidity constraints, imperfect information, and replication frictions, particularly pronounced in thematic ETFs with complex or less transparent portfolios. Thus, our results contribute to the literature by highlighting how the effectiveness of ETF arbitrage may vary systematically across investment themes, especially those characterized by innovation and heterogeneous underlying assets such as in the artificial intelligence business.

Third, we incorporate behavioral finance explanations to account for demand-driven pricing effects. AI ETFs represent a prominent example of narrative-driven investing, where strong technological optimism, media attention, and investor sentiment can generate persistent demand pressures. Behavioral theories suggest that such attention-driven flows may be relatively price-insensitive in the short run, leading to temporary mispricing and return predictability. Our findings of persistent premiums and certain return dynamics are consistent with this interpretation, indicating that thematic hype and sentiment may play a non-trivial role in the pricing of AI-related ETFs. Importantly, this behavioral channel complements, rather than contradicts, the EMH, as it operates within a framework where arbitrage is limited.

Finally, we situate our results within asset pricing theory, with particular emphasis on the distinction between systematic factor exposure and true alpha. By applying the models of Fama and French and Carhart, we show that a substantial proportion of AI ETF returns can be explained by established risk factors, including growth, profitability, investment, and momentum. This finding suggests that what may appear as “AI-driven” performance is, to a large extent, attributable to conventional factor loadings rather than to a distinct source of thematic risk premia. Consequently, our results challenge the notion that AI ETFs provide unique alpha and instead position them within the broader factor investing paradigm.

By integrating these four basic theoretical frameworks, this study moves beyond a purely empirical contribution and provides a unified interpretation of AI ETF performance. First, it is shown that return patterns are broadly consistent with conditional market efficiency. Then, it is revealed that pricing deviations can be understood through limits to ETF arbitrage mechanisms, while investor sentiment and thematic narratives contribute to demand-driven mispricing. Finally, it is accentuated that observed returns largely reflect exposure to well-established risk factors rather than novel sources of alpha. In doing so, the paper contributes to a more comprehensive theoretical understanding of how innovation-driven thematic investments are priced in modern financial markets.

3. Literature Review

Several studies in the literature have focused on AI ETFs, examining issues that mainly concern the effect of these ETFs on their underlying stocks, the financial opportunities provided by investing in AI financial instruments, the evaluation of AI ETFs and mutual funds’ performance, and the benefits from integrating AI tools, such as machine learning technology, into portfolio management.

Trying to identify the impact on underlying stocks, Wu and Chen (2022) assessed the difference in abnormal returns of underlying stocks on the inception dates of AI ETFs by classifying them into funds with AI names and those without AI names. The results show that the underlying stocks of ETFs with AI names have larger cumulative abnormal returns over the event period than stocks selected by ETFs without AI names. The difference in cumulative abnormal returns equals 40 bps. This finding verifies the existence of a return premium associated with the name of ETFs. On the other hand, Bonaparte (2024), by examining the financial opportunities arising from the new AI innovation and introducing a valuation model for AI stocks and ETFs, reported that the AI technology significantly affects financial markets, driving revenue growth and, consequently, impacting the valuations of stocks and ETFs.

On the performance of AI funds, Chen and Ren (2022) focused on the AI-powered mutual funds finding that these funds do not outperform the market. However, by comparing AI-powered funds to non-AI-powered funds, the authors reveal that the integration of AI tools into portfolio management provides a performance advantage to the corresponding funds. The authors conclude that the outperformance of AI funds is due to their lower transaction cost, superior stock-picking capability, and reduced behavioral biases. In the same field, Day and Lin (2019) developed a robo-advisor with different machine learning and deep learning forecasting methodologies to forecast ETF trends in the stock markets of Taiwan and the U.S., highlighting the superior performance achieved through AI applications compared to the traditional human-related selection process. Grobys et al. (2022) reported similar results using a sample of hedge funds. The authors point out that a strategy that involves longing hedge funds with highest level of automation and shorting funds with highest level of human involvement yields a highly significant return spread of at least 50 bps per month.

Zhang et al. (2023) examined the impact of robo-advisors on the returns of 798 Chinese ETFs, finding that the robo-advisor can enhance ETF returns up to a certain point, also revealing that the influence of a robot advisor on the average return of ETFs is comparatively less important than that of other risk and return factors. The most crucial factors are the risk attributes of ETFs. Baek et al. (2020) examined an application of machine learning to ETFs in the U.S. market by focusing on how changes in ETF prices are associated with expected market fundamentals, trying to detect long or short investment signals. The authors found that the high probability of an upward momentum suggests a long ETF signal, whereas the low probability of a downward momentum indicates a short ETF signal. Kovvuri et al. (2023) applied explainable artificial intelligence to global equity funds in the context of the G10 countries, seeking to define the relationship between diversification and performance for these funds. The results show that both over- and under-diversification are associated with poor performance.

Malhotra (2023) compared the monthly risk-adjusted returns of tech mutual funds and ETFs with the DJIA US Technology Index, the NASDAQ 100 Tech Index, the NYSE Arca Tech 100 Index, and the S&P 500 Information Technology Index over a period spanning from January 2010 to July 2021. The results indicate high correlation of these funds with the benchmark indexes. In addition, it was found that the technology mutual funds outperform ETFs and benchmarks in absolute and risk-adjusted performance. In the same context, Malhotra (2025) evaluated the performance and portfolio role of AI and Blockchain ETFs during the period 2010–2025. The findings show that both AI and Blockchain ETFs achieve positive alpha and high standalone returns, but also display significant downside risk. Moreover, the two ETF types are not significantly correlated to each other, or with broad market indices, offering potential for material diversification benefits. Karoui et al. (2024) also examined the diversification benefits and the return–risk attributes of portfolios including AI stocks. By exploring the influence of risk management strategies, ranging from “buy and hold” to daily rebalancing during the period from 30 April 2021 to 15 September 2023, the results show that the AI-related stocks have outperformed in recent years, possibly signaling an “AI bubble”.

Poutachidou and Koulis (2025) investigated the performance of ETFs that invest in AI stocks using a sample of 15 American AI ETFs over the period from 1 February 2019 to 29 December 2023. The authors focus on the degree of active versus passive management strategies adopted by AI ETFs. Based on empirical results, the AI ETFs adopt highly similar investment styles, suggesting a relatively homogeneous approach to capturing AI-related market opportunities. The results also show that most of performance variation can be explained by factor exposure and asset selection, rather than by distinct or innovative active strategies. In fact, the impact of active management on excess returns is modest and statistically insignificant, suggesting that passive investment approaches may be equally or more effective in capturing returns within the AI thematic universe. Boido and Aliano (2025) also focused on the possible benefits for AI ETFs that pursue active management with a sample of 94 AI ETFs from January 2015 to September 2023. The analysis across the entire period and sample revealed no clear advantage of active versus passive management neither an opposite advantage.

Belhouichet et al. (2025) assessed performance persistence of AI ETFs traded in the U.S. over the period 1 January 2023–23 June 2025 in comparison to other assets including the price of West Texas Intermediate (WTI) crude oil, Bitcoin, S&P 500 Index, 10-year US Treasury bonds, and the VIX Volatility Index. The results reveal a high degree of persistence for all return series. These findings suggest that the newly developed AI- and robotics-themed ETFs do not provide investors with additional hedging or diversification benefits compared to more traditional assets. Baldan et al. (2025) compared portfolios composed of AI ETFs and portfolios of core equity ETFs, utilizing the mean-variance models of Markowitz and finding that AI ETFs are superior to core equity ETFs in performance terms. Nonetheless, the outperformance of AI ETFs comes at the expense of increased uncertainty for the outcomes, which can be exacerbated in cases of AI inefficiencies or systemic risk spreading across the financial sector.

Finally, with respect to risk, Duppati et al. (2023) investigated if liquidity risk is priced in AI ETF equity returns and equity premiums, by also focusing on whether such relationships hold in extreme market situations, such as the pre- and post-COVID-19 periods. The authors found that the liquidity risk is an important determinant of AI ETFs’ valuations and equity premiums. It was also found that the equity premium tends to increase at the higher levels of bid–ask spread in the pre- and post-COVID-19 periods. Finally, Deng (2025) examined whether AI-driven funds exhibit herding behavior and how they perform under different market conditions. The key findings show that the AI funds do engage in herding, as a result of “informational herding”. Furthermore, a critical “AI gap” is revealed during bear markets, that is, AI funds significantly underperform institutional investors due to strategic inflexibility, which prevents them from adapting herding strategies to market downturns.

4. Development of Research Hypotheses

From the literature review above it is concluded that a study on the possible pricing discrepancies of AI ETFs and their impact on ETFs’ performance is missing. The respective impact of intraday volatility has also not been examined. An examination of these relationships from an “active versus passive management” research angle is missing too. Our study seeks to fill these gaps in the literature and provide reliable statistical evidence on whether the differences in trade prices and Net Asset Values of AI ETFs, as well as intraday volatility, can be beneficial to investors exposed to AI ETFs.

Based on this analysis, the research objectives of the current study can be specified via the following research hypotheses. More specifically, based on the relevant literature concerning active vs. passive management in the ETF industry, which shows that, generally, active ETFs do not outperform the market or their passive peers (e.g., Shi & Chawkat, 2024), the first hypothesis examines whether active AI ETFs outperform passive ETFs. This hypothesis is stated as follows:

H1₀. 

Active AI ETFs outperform passive AI ETFs.

H1₁. 

Active AI ETFs do not outperform passive AI ETFs.

Moreover, given that active ETFs are often considered to be riskier than passive ETFs, especially during turbulent markets, the second hypothesis to be examined regards the risk of active and passive AI ETFs. This hypothesis is stated in the following way:

H2₀. 

Active AI ETFs are riskier than passive AI ETFs.

H2₁. 

Active AI ETFs are not riskier than passive AI ETFs.

When it comes to pricing efficiency, this study examines whether there are any material differences in premiums or discounts between active and passive AI ETFs by assuming that active ETFs should present larger and more persistent pricing discrepancies than passive AI ETFs due to frictions in their arbitrage mechanism.6 The hypothesis that will be examined with respect to pricing efficiency is as follows:

H3₀. 

The pricing efficiency of active AI ETFs is inferior to that of passive AI ETFs.

H3₁. 

The pricing efficiency of active AI ETFs is not inferior to that of passive AI ETFs.

The next research issue that will be addressed regards the impact on active and passive AI ETF returns by factors including lagged returns, concurrent and lagged premiums, and concurrent and lagged intraday volatility. By discriminating between the impact on active and the impact on passive ETFs, the relevant research hypothesis can be clarified in the following way:

H4₀. 

Lagged returns, concurrent and lagged premiums, and concurrent and lagged intraday volatility affect the return of active and passive AI ETFs in the same way.

H4₁. 

Lagged returns, concurrent and lagged premiums, and concurrent and lagged intraday volatility do not affect the return of active and passive AI ETFs in the same way.

The last research objective of this study concerns the ability of active and passive AI ETFs to achieve above-market returns, given that active AI ETFs seek to outperform the market, while passive AI ETFs try to replicate market performance. The hypothesis that will be tested here is the following:

H5₀. 

Active AI ETFs outperform the market and their passive peers.

H5₁. 

Active AI ETFs do not outperform the market and their passive peers.

5. Methodology

In this section, we discuss the methodology that will be used to examine the performance of AI ETFs, the existence of pricing inefficiencies, the impact of such inefficiencies on AI ETFs’ performance, along with the impact of other ETF-related or market factors. The sample of the study is described in this section too.

5.1. Return and Risk Analysis

We calculate the return and risk of AI ETFs in daily percentage terms as follows:

$${\mathrm{R}}_{\mathrm{i}}={\displaystyle \frac{V_{i,t}-V_{i, t-1}}{V_{i,t-1}}}\times 100$$

(1)

where R_i refers to the percentage return of the ETF i on day t and V_i,t refers to the close trade price (or the NAV) of this ETF or the same day. We should highlight here that when calculating the performance of active and passive AI ETFs, the expenses charged by these funds are taken into consideration only implicitly, given that the reported Net Asset Values are computed after the managerial expenses captured by management fees. However, transactions costs are not considered in our analysis whatsoever.

Next, we compute the risk of AI ETFs as the standard deviation of daily returns. Risk is measured via the following formulas:

$${\mathsf{\sigma }}^2={\displaystyle \frac{{\sum }_{t=1}^{{\rm N}}(A_i-\overline{A}{)}^2}{{\rm N}-1}}$$

(2)

$$\mathrm{a}\mathrm{n}\mathrm{d}{\mathsf{\sigma }}_{\mathsf{A}}=\sqrt{{{\sigma }_A}^2}$$

(3)

where σ² denotes the variance of AI ETFs’ return around the average return $\overline{\mathrm{A}}$ and σ_A expresses the risk of an AI ETF portfolio in terms of standard deviation. We also calculate the risk–return (risk–reward) ratio of AI ETFs via dividing risk by average return. In general, this ratio enables evaluating potential returns against risks when deciding on investments. A lower risk–reward ratio indicates a more favorable balance between potential gains and risks, while a higher ratio suggests increased risk relative to expected return.

Along with the risk (or volatility) expressed in standard deviation terms, we calculate intraday volatility with the following formula:

$${\mathrm{I}\mathrm{n}\mathrm{t}\mathrm{V}\mathrm{o}\mathrm{l}}_{\mathrm{i}}={\displaystyle \frac{{HP}_{i,t}-{LP}_{i, t-1}}{{CP}_{i,t-1}}}\ast 100$$

(4)

where IntVol_i refers to the intraday volatility of the ETF i on day t, HP_i,t concerns the highest intraday price of this ETF on day t, LP_i,t regards the corresponding lowest intraday price, and CP_i,t denotes the close trade price of this ETF on the same day.

5.2. Premium Analysis

In this section we calculate the daily percentage difference between the close trade prices of AI ETFs and their Net Asset Values. Negative percentage differences are called “discounts” to NAV and positive differences are called “premium”.7 Premium is computed as follows:

$$Premi ={\displaystyle \frac{{CP}_{i,t}-NA{V}_{i, t}}{{NAV}_{i,t}}}\ast 100$$

(5)

where CP_i,t is as in Equation (4) and NAV_i,t is the Net Asset Value of the AI ETF i on day t.

Apart from calculating the premium of AI ETFs, we assess the persistence in daily premium via the following autoregressive model:

$${Prem}_{i,t}={\lambda }_{i,t}+{\sum }_{s=-1}^{s=-5}{Prem}_{i,s}+u_{i,t}$$

(6)

where Prem_i,t is defined as in Equation (6). The model uses five lags of daily premium as independent variables. Early findings of the ETF literature (e.g., Elton et al., 2002; Rompotis, 2009, 2010) report no premium persistence due to the unique “in-kind” creation and redemption mechanism of ETF shares, which helps eliminate significant and long-lasting differences between trade prices and NAVs. However, other studies (e.g., Hughen, 2003; Jares & Lavin, 2004; Delcoure & Zhong, 2007) find evidence of premium persistence, mostly in the case of internationally oriented ETFs. Based on this evidence, both scenarios of premium’s persistence and premium’s lack of persistence are possible.

5.3. Regression Analysis of Returns

In this section, we assess the impact of premium and intraday volatility of AI ETFs on their returns. We do so via the following time-series regression model:

R_i,t = λ_i + λ_1,iR_i,t−1 + λ_2,iPrem_i,t + λ_3,iPrem_i,t−1 + λ_4,iIntVol_i,t + λ_5,iIntVol_i,t−1 + u_i

(7)

where R_i and Prem_i are as above. IntVol_i is the intraday volatility of the AI ETF on day t.

The one-lag value of return has been added to the model for control purposes. Furthermore, if the Efficient Market Hypothesis applies to AI ETFs, there will not be any consistent relationship between return and premiums and, consequently, premium cannot be indicative of future returns. However, Jares and Lavin (2004) found a strong and significantly negative relationship between return and contemporaneous premium and a significantly positive correlation between return and lagged premium. On the contrary, Cherry (2004) finds that the return of iShares (leaders in the general ETF market) is negatively related to the one-lag discount. Rompotis (2009, 2010) reports that, on average, the return of ETFs is positively related to contemporaneous premium and negatively to lagged premium. Overall, previous findings suggest that premiums may be significant in determining the future returns of AI ETFs and that the prices of these funds may not be informationally efficient.

As noted by Shum et al. (2016), high intraday volatility in ETFs induces price swings, creating both risks, relating to magnified losses (especially for leveraged products), and opportunities (potentially significant profits from market timing/momentum strategies), often driven by volatility in the underlying assets. Based on this reasoning, we should expect the contemporaneous and the one-lag intraday volatility factors in Model (7) to be quite significant in explaining the return of AI ETFs.

5.4. Performance Regression Analysis

In the previous section, we tried to identify how the return of active and passive AI ETFs is affected by lagged return, concurrent and lagged pricing inefficiencies, and concurrent and lagged intraday volatility. In this section, we analyze the performance of AI ETFs by incorporating traditional risk factors included in the Fama and French (2015) five-factor model, as well as Fama and French’s version of Carhart’s (1997) momentum factor.

The five-factor model provides a broader risk–return framework, trying to separate true abnormal performance from returns explained by exposure to risk factors. At the same time, momentum captures short-term performance trends. By decomposing returns into contributions from Fama and French’s five factors and the momentum factor of Carhart, investors can better understand the sources of excess returns of AI ETFs and gauge risk-adjusted performance, determining if the excess returns achieved by AI ETFs can justify the risks taken.

Each factor in the model represents a systematic source of risk or behavioral return pattern. More specifically, this model assesses the ability of AI ETFs to achieve any material above-market return, measures their systematic risk and assesses their exposure to certain market factors such as size, value, conservativeness, robustness and momentum.

The size effect implies that small-cap stocks outperform the large ones. The book-to-market equity ratio effect captured by the HML factor shows that the average returns on stocks with high book-value to market-value equity ratios is greater than the returns on stocks with low such ratios. The Conservative Minus Aggressive and Robust Minus Weak factors correspond to the Fama and French (2015) investment and operating profitability factors. The authors consider past investment as a proxy for the expected future investment and suggest that CMA implies a negative relationship between the expected investment and the expected internal rate of return, while they reveal a negative loading for the RMW factor. Finally, the existence of momentum in asset prices is an anomaly that is difficult to explain given that, as the efficient capital markets theory suggests, an increase in the price of an asset cannot be indicative of further future increases in prices. The time series model applied is the following:

R_i − R_f = α_i + β_1,i(R_m − R_f) + β_2,iSMB + β_3,iHML + β_4,iCMA + β_5,iRMW + β_6,iMOM + u_i

(8)

where R_i is defined as in Formula (1). R_m is the daily return of the market portfolio, which is proxied by the Russell 3000 Index.8 It is worth noticing that quite often, AI ETFs have high market beta due to their concentration in tech equities. R_f is the risk-free rate expressed by the one-month U.S. Treasury bill rate. The SMB and HML factors in Model (8) have been suggested by Fama and French (1993). SMB (Small Minus Big) is the average return on nine small-cap portfolios minus the average return on nine big-cap portfolios. This factor assesses whether the returns of AI ETFs come from smaller innovative firms rather than from large- or mega-cap tech companies. HML (High Minus Low) is the average return on two value portfolios (based on book-to-market equity values) minus the average return on two growth portfolios. AI ETFs tend to have negative HML exposure because they hold growth-oriented tech firms.

The CMA and RMW factors have been added by Fama and French (2015) to their original model of 1993. CMA (Conservative Minus Aggressive) is the average return on two conservative portfolios minus the average return on two aggressive portfolios. The tendency of AI companies to reinvest heavily in Research and Development (aggressive investment) should lead to negative CMA exposure for the corresponding AI ETFs. RMW (Robust Minus Weak) is the average return on two robust operating profitability portfolios minus the average return on two weak operating profitability portfolios. As AI ETFs hold usually profitable tech leaders, they show a positive RMW loading.

Finally, MOM (momentum) is the average of the returns on two (big and small) high prior return portfolios minus the average of the returns on two low prior return portfolios.9 AI stocks often display strong momentum effects relating to the enthusiasm of investors that drives momentum trends, the price continuation supported by rapid technological narratives, and the inflows from institutional investors who chase high-growth tech sectors. The inclusion of the momentum factor in the performance regression model explores whether any excess return generated by an AI ETF can be attributed to momentum exposure rather than managerial skill. That said, if the momentum factor is omitted from the model, AI ETFs may show positive alphas, but once momentum is included, the alpha may disappear because performance is actually due to trend-following exposure (i.e., false alphas). As Malhotra (2025) notes, the inclusion of the momentum factor in the performance regression analysis is especially relevant for AI ETFs, as the stocks underneath these funds often experience sharp valuation swings affected by technological breakthroughs, innovation cycles, and investor sentiment.

Overall, Fama and French’s five factors capture structural firm characteristics considering size, value, profitability, and investment, while momentum captures behavioral market dynamics. AI ETFs typically show negative HML exposure, negative CMA loading and positive momentum. This combination often characterizes innovation-driven technology portfolios. In our case, if the exposures to the six factors capture all variation in the expected returns of AI ETFs, alphas will be zero. However, Malhotra (2025) finds that the AI (and Blockchain) ETFs do generate positive alphas via applying similar regression analysis (with monthly returns instead of daily returns used in our analysis). Thus, if Malhotra’s results hold in our sample too, positive and significant alphas are to be expected.

5.5. The Sample

According to Etf.com, currently there are 81 ETFs in the U.S. that are classified as AI-related. The total assets under management of these ETFs amount to $34.98B. The majority of these AI ETFs are actively managed (i.e., 41 funds), 27 of them pursue passive management and 13 are leveraged or inverse leveraged ETFs.

In assembling the sample of our study, we required that an AI ETF has full trading data for at least one year up to 31 December 2025. In addition, we decided to exclude the leveraged and inverse leveraged ETFs from our analysis due to their unique investing orientation, which adopts the target of achieving X times the performance of a benchmark in a positive or negative way, compared to passive ETFs that simply seek to replicate their benchmarks or the active ETFs that aim at beating the overall market. Based on our selection criteria, 25 active and 22 passive AI ETFs made the cut to be included in our sample.

At this point, a comment that needs to be made is that our sample does not suffer from any material survivorship bias, given that AI ETFs are a relatively new entry in the ETF market (as inferred by the average age of these funds below). In fact, according to Nasdaq.com, most major AI ETFs launched since 2016 to 2023 are still active, benefiting from strong investor demand post-ChatGPT boom.10 This element suggests that our results are free from any survivorship bias.

Table 1. Profiles of ETFs.

Symbol	Name	Type	Launch Date	Age	Assets ($Mils)	Exp. Ratio	Volume (Shares)	ESG Score
Panel A: Active ETFs
BAI	iShares A.I. Innovation and Tech Active ETF	Active	21 October 2024	1.19	8210	0.55	1,256,185	6.30
ARKQ	ARK Autonomous Technology & Robotics ETF	Active	30 September 2014	11.26	2000	0.75	146,214	6.20
IETC	iShares U.S. Tech Independence Focused ETF	Active	21 March 2018	7.79	945	0.18	24,006	6.57
AIVL	WisdomTree U.S. AI Enhanced Value Fund	Active	16 June 2006	19.56	384	0.38	42,023	6.59
AIVI	WisdomTree International AI Enhanced Value Fund	Active	16 June 2006	19.56	58	0.58	27,037	7.62
IEDI	iShares U.S. Consumer Focused ETF	Active	21 March 2018	7.79	29	0.18	4812	6.34
AMOM	QRAFT AI-Enhanced U.S. Large Cap Momentum ETF	Active	20 May 2019	6.62	26	0.75	6155	6.05
QRFT	QRAFT AI-Enhanced U.S. Large Cap ETF	Active	20 May 2019	6.62	17	0.75	2691	6.25
CHAT	Roundhill Generative AI & Technology ETF	Active	18 May 2023	2.62	1020	0.75	136,330	6.29
AIPI	REX AI Equity Premium Income ETF	Active	4 June 2024	1.58	409	0.65	118,897	6.58
ALAI	Alger AI Enablers & Adopters ETF	Active	5 April 2024	1.74	295	0.55	39,677	5.57
FDTX	Fidelity Disruptive Technology ETF	Active	12 June 2023	2.56	199	0.50	21,947	6.82
FBOT	Fidelity Disruptive Automation ETF	Active	12 June 2023	2.56	174	0.50	16,028	6.99
ECML	Euclidean Fundamental Value ETF	Active	18 May 2023	2.62	136	0.95	3649	5.87
AGIX	KraneShares Artificial Intelligence & Technology ETF	Active	18 July 2024	1.45	123	0.99	19,618	6.32
AIS	VistaShares Artificial Intelligence Supercycle ETF	Active	3 December 2024	1.08	122	0.75	34,128	6.73
AIFD	TCW Artificial Intelligence ETF	Active	6 May 2024	1.65	86	0.75	12,253	6.75
GROZ	Zacks Focus Growth ETF	Active	5 December 2024	1.07	56	0.56	12,337	6.12
AIYY	YieldMax AI Option Income Strategy ETF	Active	27 November 2023	2.10	55	1.67	67,392	5.77
LRNZ	TrueShares Technology, AI & Deep Learning ETF	Active	28 February 2020	5.84	33	0.69	8104	6.70
DRAI	Draco Evolution AI ETF	Active	10 July 2024	1.48	22	1.34	4725	5.92
GPT	Intelligent Alpha Atlas ETF	Active	17 September 2024	1.29	21	0.69	9416	6.62
NEWZ	StockSnips AI-Powered Sentiment US All Cap ETF	Active	12 April 2024	1.72	19	0.75	4544	6.52
XDAT	Franklin Exponential Data ETF	Active	12 January 2021	4.97	4	0.50	1496	5.89
LQAI	LG QRAFT AI-Powered U.S. Large Cap Core ETF	Active	7 November 2023	2.15	3	0.75	1090	6.05
Mean				4.75	578	0.70	80,830	6.38
Min				1.07	3	0.18	1090	5.57
Max				19.56	8210	1.67	1,256,185	7.62
Panel B: Passive ETFs
AIQ	Global X Artificial Intelligence & Technology ETF	Passive	11 May 2018	7.65	7700	0.68	269,051	6.79
BOTZ	Global X Robotics & Artificial Intelligence ETF	Passive	12 September 2016	9.31	3330	0.68	769,925	7.04
ARTY	iShares Future AI & Tech ETF	Passive	26 June 2018	7.52	2080	0.47	113,206	6.35
ROBO	ROBO Global Robotics & Automation Index ETF	Passive	22 October 2013	12.20	1380	0.95	145,671	6.58
ROBT	First Trust Nasdaq Artificial Intelligence & Robotics E	Passive	21 February 2018	7.86	714	0.65	41,136	6.80
IGPT	Invesco AI and Next Gen Software ETF	Passive	23 June 2005	20.54	677	0.56	64,391	6.67
DRIV	Global X Autonomous & Electric Vehicles ETF	Passive	13 April 2018	7.72	350	0.68	166,202	6.30
AIEQ	Amplify AI Powered Equity ETF	Passive	17 October 2017	8.21	118	0.75	32,166	6.23
AIVC	Amplify Bloomberg AI Value Chain ETF	Passive	8 March 2016	9.82	35	0.59	5044	7.02
WTAI	WisdomTree Artificial Intelligence and Innovation Fun	Passive	7 December 2021	4.07	437	0.45	99,724	6.78
THNQ	ROBO Global Artificial Intelligence ETF	Passive	11 May 2020	5.64	298	0.68	15,448	6.64
BUZZ	VanEck Social Sentiment ETF	Passive	2 March 2021	4.84	112	0.76	77,276	4.99
XAIX	Xtrackers Artificial Intelligence and Big Data ETF	Passive	2 Aug 2024	1.41	102	0.35	21,284	6.84
RWEM	Rayliant Wilshire NxtGen Emerging Markets Equity E	Passive	16 December 2021	4.04	83	0.52	7121	6.25
RWLC	Rayliant Wilshire NxtGen US Large Cap Equity ETF	Passive	16 December 2021	4.04	83	0.32	10,690	6.61
HEAL	Global X HealthTech ETF	Passive	29 July 2020	5.43	43	0.50	36,044	5.95
WISE	Themes Generative Artificial Intelligence ETF	Passive	8 December 2023	2.07	39	0.35	11,012	5.37
IBOT	Vaneck Robotics ETF	Passive	5 April 2023	2.74	34	0.47	4213	7.55
FAI	First Trust Bloomberg Artificial Intelligence ETF	Passive	20 November 2024	1.11	33	0.65	6382	6.81
BOTT	Themes Humanoid Robotics ETF	Passive	22 April 2024	1.69	19	0.35	2279	N/A
KCAI	KraneShares China Alpha Index ETF	Passive	28 Aug 2024	1.34	8	0.79	2336	4.17
BULD	Pacer BlueStar Engineering the Future ETF	Passive	4 May 2022	3.66	2	0.49	374	7.85
Mean				6.04	804	0.58	86,408	6.46
Min				1.11	2	0.32	374	4.17
Max				20.54	7700	0.95	769,925	7.85

This table presents the profiles of U.S.-listed AI ETFs examined in this study. Profiles include the symbol and name of ETFs, type, launch date, age, assets ($Millions) as at 31 December 2025, expense ratio, average daily volume (in shares) over the entire trade history of each ETF and ESG score.

contains information on the sample AI ETFs. Information provided concerns the symbol and name of ETFs, type (i.e., active or passive), launch date, age, assets as at 31 December 2025, expense ratio, average daily volume (in shares) over the entire trade history of each ETF and ESG score.

Passive AI ETFs are older than their active peers by about 1.3 years. The average passive AI ETF is about 6 years old, while the average active AI ETF is about 4.7 years old. The oldest AI ETF in the sample is the Invesco AI and Next Gen Software ETF (IGPT). The average active AI ETF held about $578M as at 31 December 2025. The respective figure for passive ETFs is equal to $804M. The largest ETF in the entire sample is the iShares A.I. Innovation and Tech Active ETF (BAI), with assets of $8.2B. Overall, the total assets of the 47 AI ETFs included in our sample at the end of 2025 amounted to $32.1B, that is, about 92% of total assets channeled to all AI ETFs. This percentage indicates that the sample of our study is quite representative of the AI niche of the ETF market in the United States.

Furthermore, the average expense ratio of active AI ETFs is equal to 0.70%, being larger than that of passive AI ETFs by 12 bps, which is equal to 0.58%. Notably, the largest expense ratio in the sample is equal to 1.67%, charged by the active YieldMax AI Option Income Strategy ETF. With respect to trading activity, the average daily volume of active AI ETFs amounts to 81 thousand shares. The corresponding average volume for passive ETFs is slightly higher at 86 thousand shares. Finally, on the question of the ESG performance of the AI ETFs examined, the figures in Table 1 are rather modest. Active ETFs have an average ESG scores of 6.38, whereas the average ESG scores of passive AI ETFs is equal to 6.46.

6. Empirical Results

6.1. Return and Risk Analysis

Table 2. Return and risk analysis.

Symbol	Price Ret	NAV Ret	Price Cum	NAV Cum	Price Ret Min	NAV Ret Min	Price Ret Max	NAV Ret Max	Price Risk	NAV Risk	Price Risk/ PriceRet	NAVRisk/ NAVRet	IntVol
Panel A: Active ETFs
BAI	0.12	0.13	35.48	35.92	−7.75	−7.79	13.55	13.61	2.13	2.15	17.05	17.02	2.14
ARKQ	0.08	0.08	489.33	487.86	−10.44	−11.67	14.05	13.96	1.79	1.80	22.71	22.83	1.78
IETC	0.09	0.09	323.21	314.22	−13.26	−12.66	12.41	12.54	1.61	1.62	18.49	18.87	1.44
AIVL	0.02	0.02	131.30	131.18	−11.89	−11.74	11.37	11.60	1.22	1.25	49.77	50.14	1.08
AIVI	0.01	0.01	4.37	4.47	−13.18	−12.02	15.76	11.12	1.40	1.22	130.60	145.57	0.98
IEDI	0.05	0.05	124.90	122.49	−11.90	−12.09	8.18	8.11	1.24	1.25	25.21	25.61	0.76
AMOM	0.05	0.05	93.50	93.98	−19.39	−19.42	10.36	10.40	1.66	1.67	30.87	30.87	1.09
QRFT	0.06	0.06	149.30	149.31	−9.94	−11.55	9.62	9.65	1.29	1.32	20.42	20.73	0.67
CHAT	0.14	0.14	127.58	128.34	−8.94	−8.17	14.51	11.39	1.81	1.61	12.77	11.63	1.81
AIPI	−0.05	−0.05	−20.96	−20.95	−7.01	−6.93	10.23	10.41	1.49	1.49	−30.72	−30.89	1.44
ALAI	0.15	0.15	79.79	80.42	−7.43	−7.25	14.16	12.96	1.81	1.77	11.97	11.73	1.42
FDTX	0.09	0.08	59.66	59.96	−6.84	−6.44	11.95	11.94	1.56	1.49	18.36	17.67	1.52
FBOT	0.05	0.05	30.85	31.36	−6.42	−5.63	9.91	8.71	1.29	1.09	25.63	22.50	1.10
ECML	0.05	0.05	31.83	32.00	−6.65	−6.09	7.42	7.24	1.20	1.20	24.34	24.31	0.37
AGIX	0.12	0.12	46.39	46.61	−7.94	−7.03	11.95	11.64	1.83	1.75	15.13	14.57	1.72
AIS	0.18	0.17	50.56	50.51	−9.38	−8.51	13.61	9.19	2.19	1.88	12.42	11.06	2.10
AIFD	0.11	0.11	47.10	46.94	−7.60	−7.64	13.43	13.62	1.85	1.85	16.82	16.87	1.61
GROZ	0.07	0.07	18.07	18.24	−6.13	−6.06	11.66	11.25	1.50	1.49	20.41	20.16	1.17
AIYY	−0.41	−0.41	−91.38	−91.37	−24.36	−24.86	16.03	19.35	3.33	3.46	−8.12	−8.53	3.65
LRNZ	0.07	0.07	87.93	88.11	−11.93	−13.31	14.54	14.33	2.39	2.42	33.38	33.43	2.10
DRAI	0.05	0.05	18.99	19.05	−5.05	−5.45	5.56	5.74	1.09	1.10	20.71	20.74	0.58
GPT	0.06	0.06	19.46	19.61	−7.94	−8.04	9.57	9.51	1.29	1.30	20.26	20.29	1.05
NEWZ	0.03	0.03	9.94	10.01	−6.11	−6.21	5.40	5.46	1.03	1.02	37.80	37.21	0.51
XDAT	0.02	0.02	4.03	4.79	−8.63	−7.84	11.76	11.63	1.88	1.87	90.17	88.06	0.79
LQAI	0.09	0.09	55.27	55.27	−5.91	−5.83	9.70	9.55	1.08	1.08	12.29	12.29	0.32
Mean	0.05	0.05	77.06	76.73	−9.68	−9.61	11.47	11.00	1.64	1.61	25.95	26.19	1.33
Min	−0.41	−0.41	−91.38	−91.37	−24.36	−24.86	5.40	5.46	1.03	1.02	−30.72	−30.89	0.32
Max	0.18	0.17	489.33	487.86	−5.05	−5.45	16.03	19.35	3.33	3.46	130.60	145.57	3.65
Panel B: Passive ETFs
AIQ	0.08	0.08	242.03	242.00	−9.83	−11.36	12.19	9.95	1.60	1.49	20.77	19.79	1.57
BOTZ	0.05	0.05	145.46	145.39	−12.41	−8.53	12.77	9.20	1.62	1.31	31.39	27.82	1.46
ARTY	0.05	0.05	104.07	104.65	−10.22	−9.52	13.18	13.66	1.71	1.65	32.57	31.94	1.70
ROBO	0.04	0.04	173.74	176.27	−11.16	−8.97	10.23	8.50	1.36	1.13	32.31	28.63	1.29
ROBT	0.04	0.04	72.83	73.22	−11.86	−11.32	11.64	11.50	1.60	1.57	39.48	39.09	1.50
IGPT	0.06	0.06	1104.66	1109.94	−11.96	−12.53	13.36	13.86	1.50	1.52	25.39	25.49	1.52
DRIV	0.05	0.05	93.61	94.87	−11.98	−11.85	11.21	10.46	1.72	1.58	35.15	33.60	1.67
AIEQ	0.04	0.04	78.93	79.00	−12.00	−12.00	11.87	12.33	1.46	1.49	37.54	37.87	1.44
AIVC	0.05	0.05	168.01	168.61	−11.62	−11.62	14.00	10.04	1.65	1.52	30.81	29.44	1.41
WTAI	0.03	0.03	17.45	17.67	−7.38	−7.23	13.09	12.65	1.91	1.78	56.14	55.98	1.63
THNQ	0.08	0.08	151.00	151.72	−6.80	−6.81	12.78	11.79	1.79	1.74	22.15	21.66	1.75
BUZZ	0.05	0.05	38.14	38.16	−7.11	−7.16	14.44	14.79	2.05	2.05	43.07	43.04	2.22
XAIX	0.12	0.12	48.02	48.18	−6.12	−5.75	11.47	9.56	1.41	1.29	11.69	10.83	1.28
RWEM	0.02	0.02	17.67	16.82	−6.32	−6.02	8.27	4.50	1.19	0.95	51.51	47.90	0.85
RWLC	0.03	0.03	32.64	33.56	−11.71	−12.94	8.79	6.31	1.11	1.00	32.49	29.74	0.83
HEAL	−0.02	−0.02	−37.45	−36.87	−6.09	−5.90	8.03	7.01	1.66	1.56	−80.33	−72.43	1.92
WISE	0.10	0.10	53.39	53.27	−6.89	−6.23	13.94	12.67	2.09	1.99	19.99	19.43	1.96
IBOT	0.07	0.07	54.35	54.15	−6.56	−6.68	11.45	12.04	1.35	1.36	18.67	18.81	0.85
FAI	0.13	0.13	36.07	36.41	−7.18	−7.09	14.36	13.87	1.92	1.88	14.77	14.44	1.32
BOTT	0.14	0.14	69.11	68.68	−7.27	−5.96	11.78	4.84	1.84	1.58	13.07	11.62	0.95
KCAI	0.08	0.08	22.51	22.35	−25.81	−25.69	6.93	6.37	2.04	1.83	24.46	23.18	0.21
BULD	0.04	0.04	30.73	30.24	−8.44	−5.70	10.61	7.54	1.73	1.55	39.19	37.94	0.36
Mean	0.06	0.06	123.50	124.01	−9.85	−9.40	11.66	10.16	1.65	1.54	25.10	24.36	1.35
Min	−0.02	−0.02	−37.45	−36.87	−25.81	−25.69	6.93	4.50	1.11	0.95	−80.33	−72.43	0.21
Max	0.14	0.14	1104.66	1109.94	−6.09	−5.70	14.44	14.79	2.09	2.05	56.14	55.98	2.22

This table presents the trading statistics of AI ETFs over their entire trade history, which include their average daily return, cumulative return, minimum and maximum returns, risk, risk to return (reward) ratio, intraday volatility calculated as the percentage difference between the intraday highest minus the intraday lowest trade price of an ETF divided by its close trade price on day t.

presents the return and risk of AI ETFs in trade price and NAV terms. Statistics reported regard average daily return, cumulative return over the entire trade history of each ETF, minimum and maximum return scores, risk estimates (i.e., standard deviation in daily returns), the risk–return ratio, and intraday volatility. With the exception of intraday volatility, which is computed with trade prices, all of the statistics are calculated both with trade prices and Net Asset Values.

The average daily return of active AI ETFs is equal to 5 bps, both in trade price and NAV terms. The corresponding return of passive ETFs is higher by 1 bps. At the cumulative level, passive AI ETFs have clearly outperformed active ETFs over their entire trade history by about 46–47% (for both types of returns considered). This finding supports the merit of passive strategies in the long-lasting debate in finance of “active vs. passive management”. At the fund level, only two active ETFs and just one passive achieved negative cumulative returns, verifying that AI ETFs are a very promising investment choice, even though high losses are possible, as evidenced by the minimum return records in Table 2.

The outperformance of passive AI ETFs carries important economic implications for both financial markets and investment practice. Persistent passive outperformance challenges the traditional rationale for active management and lends support to the Efficient Market Hypothesis, which posits that consistently generating above-market returns through security selection is inherently difficult. In the specific context of AI ETFs, the documented return superiority of passive strategies raises questions about the widespread adoption of active management within this segment and whether such approaches are justified given their performance outcomes.

From an asset pricing perspective, the outperformance of passive AI ETFs suggests that a substantial portion of returns may be attributable to systematic factor exposures embedded in passive index strategies, rather than to superior managerial skill. This interpretation is consistent with the broader evidence that factor-driven returns are difficult to outperform after accounting for fees and transaction costs.

Moreover, the growing dominance of passive ETFs may have implications for price formation. Increased demand for index constituents can generate price pressure and potentially influence cross-sectional return patterns, particularly in concentrated sectors such as artificial intelligence. At the same time, the low-cost structure and transparency of passive AI ETFs are likely to enhance investor welfare by reducing fees, improving accessibility, and facilitating efficient portfolio allocation.

Overall, these findings highlight the need to carefully evaluate the role of active management in thematic investing and to distinguish between returns driven by skill and those arising from systematic exposures.

Furthermore, with respect to volatility, the average risk of active AI ETFs equals 1.64 (1.61) in trade price (NAV) terms. The respective figures for passive ETFs are 1.65 and 1.54. The comparison of active and passive ETFs shows that the return advantage of passive ETFs over active ETFs is not attained at a cost to investors, in terms of increased risk (and increased expense ratios as shown in Table 1). This assertion is confirmed by the risk–return ratios, which are higher for active AI ETFs than passive ETFs by 85 (183) bps in trade price (NAV) terms. Overall, the combined analysis of return and risk manifests a significant advantage of passive AI ETFs against their active counterparts.

Finally, the average terms of intraday volatility for active and passive ETFs verify that the latter are not riskier than the former. The average intraday volatility of active ETFs equals 1.33 and the respective figure of passive ETFs is slightly higher (by 2 bps) at 1.35. It is worth noticing that, occasionally, intraday volatility can be either quite low or materially high, as shown by the minimum and maximum intraday volatility records for both groups. Based on these extreme records, the passively managed AI ETFs are superior to the active ETFs in volatility terms (i.e., they are less volatile intraday).

The differences in performance and risk between active and passive AI ETFs may be attributed to differences in portfolio construction, market timing and managerial skills, concentration and diversification, exposure to risk factors, cost basis, trading activity and liquidity, differences in defining the AI investment theme, and exposure to innovation risk.

Regarding portfolio construction, as well as market timing and managerial skills, passive AI ETFs track an AI-themed index and replicate its holdings (typically large, well-known technology firms). This approach prescribes fixed rules for securities’ inclusion and weights, which allow limited flexibility during market shifts and less ability to avoid bubbles or overvalued stocks, but contribute to lower tracking errors. On the other hand, active AI ETFs rely on portfolio managers to select stocks they believe will benefit most from AI development. Based on their research, the managers of active AI ETFs may choose to invest in smaller or emerging AI companies, having the ability to adjust exposure during market volatility. In other words, the managers of active AI ETFs can increase exposure during AI growth cycles and reduce holdings during tech corrections. Successful timing entails higher returns, but poor timing can lead to underperformance relative to passive AI ETF peers.

Furthermore, as passive AI ETFs are often highly concentrated in mega-cap tech firms (such as NVIDIA or Alphabet Inc.), they might be exposed to lower individual stock risk, but they depend strongly on the technology sector cycle. Active AI ETFs may invest more heavily in AI startups or niche companies and, thus, portfolio concentration can be higher for these funds. As a result, volatility can be higher for active AI ETFs than that of passive AI ETFs.

Another explanation of the differences in return and risk of active and passive AI ETFs concerns the possibly different factor exposures (such as the factors in Fama and French (1993, 2015), and the momentum factor of Carhart (1997)) between the two types of AI ETFs. Different factor exposures produce different risk–return profiles.11 When it comes to costs, active ETFs usually have higher expense ratios than passive ETFs. Higher costs reduce net returns over time. Passive ETFs often outperform their active peers after fees.

On the question of trading activity and liquidity, active AI ETFs have higher turnover as managers need to adjust portfolios more frequently than the passive managers do, entailing higher transaction costs and tax implications for active AI ETFs than the passive ones, which have lower turnover due to index tracking and more stable long-term holdings. With respect to the differences in AI thematic definition, passive indices apply strict rules to define “AI companies”, while active managers may use broader interpretations, possibly leading to different portfolios and, consequently, different performance and risk outcomes. Finally, as far as exposure to innovation risk is concerned, given that AI is a rapidly evolving industry, passive AI ETFs tend to include established tech firms with lower technological uncertainty, whereas active AI ETFs may invest earlier in disruptive AI technologies with higher growth potential, but with higher failure risk too.

6.2. Premium Analysis

The pricing of AI ETFs, i.e., the differences between trade prices and NAVs, is discussed in this section. The relevant calculations are provided in

Table 3. Premium analysis.

Symbol	Average	Min	Max	StDev	Days of Discount	Days of Premium	% of Days with Discount	% of Days with Premium	Mean Disc	Mean Prem
Panel A: Active ETFs
BAI	0.12	−0.31	2.23	0.17	47	251	15.772	84.228	−0.08	0.16
ARKQ	0.04	−1.88	2.50	0.26	1246	1584	44.028	55.972	−0.12	0.17
IETC	0.03	−2.09	1.92	0.16	712	1242	36.438	63.562	−0.08	0.10
AIVL	−0.01	−1.47	1.91	0.11	2835	2081	57.669	42.331	−0.07	0.07
AIVI	0.00	−6.87	8.51	0.73	2349	2567	47.783	52.217	−0.49	0.46
IEDI	−0.04	−3.90	2.17	0.29	1290	664	66.018	33.982	−0.13	0.12
AMOM	−0.03	−2.11	2.17	0.22	1011	653	60.757	39.243	−0.13	0.12
QRFT	−0.02	−1.95	0.98	0.18	919	745	55.228	44.772	−0.10	0.08
CHAT	0.03	−2.14	1.56	0.31	296	362	44.985	55.015	−0.22	0.24
AIPI	0.09	−0.22	0.83	0.10	68	328	17.172	82.828	−0.06	0.12
ALAI	0.08	−0.57	1.25	0.17	128	309	29.291	70.709	−0.09	0.15
FDTX	0.00	−0.92	1.15	0.19	324	318	50.467	49.533	−0.14	0.14
FBOT	−0.02	−2.14	1.40	0.44	309	333	48.131	51.869	−0.36	0.29
ECML	0.01	−0.59	0.36	0.07	283	375	43.009	56.991	−0.04	0.06
AGIX	0.12	−1.42	1.62	0.31	91	275	24.863	75.137	−0.22	0.22
AIS	0.21	−1.86	2.92	0.44	67	203	24.815	75.185	−0.29	0.38
AIFD	−0.03	−0.49	0.30	0.12	247	169	59.375	40.625	−0.11	0.09
GROZ	0.04	−0.23	1.27	0.14	104	164	38.806	61.194	−0.05	0.10
AIYY	0.09	−2.45	4.75	0.53	229	296	43.619	56.381	−0.41	0.31
LRNZ	0.02	−0.51	3.17	0.22	800	668	54.496	45.504	−0.07	0.13
DRAI	0.04	−0.19	0.54	0.08	88	284	23.656	76.344	−0.05	0.07
GPT	0.05	−0.67	0.77	0.15	131	192	40.557	59.443	−0.07	0.13
NEWZ	0.09	−0.59	1.81	0.14	63	369	14.583	85.417	−0.06	0.11
XDAT	0.02	−3.99	5.50	0.49	615	632	49.318	50.682	−0.14	0.16
LQAI	−0.01	−0.54	0.54	0.08	284	255	52.690	47.310	−0.06	0.05
Mean	0.04	−1.60	2.09	0.24	581	613	41.74	58.26	−0.15	0.16
Min	−0.04	−6.87	0.30	0.07	47	164	14.58	33.98	−0.49	0.05
Max	0.21	−0.19	8.51	0.73	2835	2567	66.02	85.42	−0.04	0.46
Panel B: Passive ETFs
AIQ	0.06	−2.40	1.53	0.29	719	1200	37.467	62.533	−0.24	0.20
BOTZ	0.06	−4.75	4.56	0.57	981	1358	41.941	58.059	−0.45	0.38
ARTY	0.14	−2.26	1.43	0.27	547	1341	28.972	71.028	−0.18	0.27
ROBO	0.00	−4.28	3.98	0.43	1458	1609	47.538	52.462	−0.67	0.29
ROBT	0.12	−3.08	2.43	0.25	488	1488	24.696	75.304	−0.17	0.22
IGPT	0.00	−4.25	4.33	0.24	2760	2403	53.457	46.543	−0.13	0.13
DRIV	−0.02	−7.71	2.92	0.37	1036	903	53.430	46.570	−0.28	0.22
AIEQ	−0.02	−0.91	1.12	0.16	1157	905	56.111	43.889	−0.13	0.11
AIVC	−0.01	−2.37	2.69	0.39	1338	1131	54.192	45.808	−0.28	0.31
WTAI	0.13	−3.21	5.54	0.53	354	665	34.740	65.260	−0.23	0.31
THNQ	0.04	−1.20	0.81	0.22	563	856	39.676	60.324	−0.17	0.18
BUZZ	−0.04	−2.92	2.50	0.16	742	472	61.120	38.880	−0.10	0.06
XAIX	0.15	−0.59	1.45	0.18	57	298	16.056	83.944	−0.11	0.20
RWEM	0.09	−6.35	7.28	0.98	457	557	45.069	54.931	−0.61	0.64
RWLC	−0.10	−3.59	4.47	0.50	644	370	63.511	36.489	−0.35	0.32
HEAL	−0.13	−1.61	1.03	0.36	878	485	64.417	35.583	−0.36	0.22
WISE	0.17	−1.40	0.93	0.23	109	408	21.083	78.917	−0.17	0.25
IBOT	0.23	−0.44	0.74	0.14	24	663	3.493	96.507	−0.10	0.25
FAI	0.16	−0.35	0.88	0.15	27	250	9.747	90.253	−0.07	0.19
BOTT	0.34	−2.17	5.52	0.59	83	343	19.484	80.516	−0.40	0.52
KCAI	0.09	−3.71	13.28	1.16	159	178	47.181	52.819	−0.36	0.48
BULD	−0.02	−9.33	6.28	1.00	418	500	45.534	54.466	−0.53	0.38
Mean	0.07	−3.13	3.44	0.42	682	836	39.50	60.50	−0.28	0.28
Min	−0.13	−9.33	0.74	0.14	24	178	3.49	35.58	−0.67	0.06
Max	0.34	−0.35	13.28	1.16	2760	2403	64.42	96.51	−0.07	0.64

This table presents the statistics of AI ETFs’ premium (discount) over their entire trade history, which include their average daily percentage premium (discount), minimum and maximum premiums (discounts), the standard deviation of premiums (discounts), the number and the portion of days with premiums and discounts, and the corresponding mean discount and premium.

. Average terms, minimum and maximum records, and standard deviations are included in the table. The number and portion (%) of days with discounts and premiums are provided too. The corresponding mean discounts and mean premiums are also presented. The regression results on premium persistence are discussed in this section too.

The average active AI ETF trades at premium of 4 bps, while the average premium of passive AI ETFs is slightly higher at 7 bps. At the fund level, 8 out of 25 active ETFs and 9 out of 22 passive ETFs trade at a discount. Furthermore, quite high daily discounts or premiums are to be expected when trading with AI ETFs, as verified by the extreme scores. The extreme discounts or premiums are higher for passive ETFs than those of active ETFs, showing that the arbitrage mechanism inherent to ETFs, which helps minimize the differences between their trade prices and NAVs, works better for active AI ETFs than for their passive peers. On the variation in daily premiums, the relevant standard deviations in Table 3 indicate that pricing discrepancies are not that volatile. However, they are more volatile for passive AI ETFs than for active ETFs (i.e., 0.42 vs. 0.24, respectively).

Overall, the analysis of pricing efficiency so far shows that active AI ETFs trade closer to their underlying values compared to passive ETFs. This evidence could be critical for investors forming their ETF investment choices based on the criterion of proximity between the value of ETF shares and the underlying fundamental values.

Furthermore, as shown in Table 3, both active and passive AI ETFs are more prone to trading at a premium instead of a discount. On about 60% of total trade history, active and passive AI ETFs present a premium at the end of the trading day, while discounts are presented on about 40% of days. In the case of active (passive) ETFs, the respective mean discount amounts to 15 (28) bps, while their mean premium is equal to 16 (28) bps. These statistics verify the superiority of active AI ETFs in trading closer to their NAVs.

The estimates of Model (6) on premium persistence are found in

Table 4. Premium persistence regression analysis.

Symbol	Const	t-Stat	Pr(-1)	t-Stat	Pr(-2)	t-Stat	Pr(-3)	t-Stat	Pr(-4)	t-Stat	Pr(-5)	t-Stat	R2	Obs
Panel A: Active ETFs
BAI	0.13 a	6.06	−0.02	−0.31	0.15 b	2.03	−0.04	−0.56	0.00	0.01	−0.10	−1.41	0.12	293
ARKQ	0.03 a	6.07	0.05	1.16	0.10 c	1.89	0.03	0.79	0.08 c	1.90	0.03	0.76	0.13	2825
IETC	0.02 a	4.61	0.07	0.68	0.01	0.27	0.13 a	2.98	0.07	1.45	−0.01	−0.08	0.13	1949
AIVL	−0.01 a	−3.35	0.08	1.19	0.13 a	3.22	0.12 a	2.68	0.05	1.12	0.01	0.30	0.15	4911
AIVI	0.00	0.28	0.03	0.86	0.09 a	2.60	0.11 a	3.43	0.07	1.61	0.08 b	2.25	0.14	4911
IEDI	−0.02 a	−3.49	0.31 a	2.62	0.08	1.28	0.05	0.62	−0.04	−0.46	0.06	0.67	0.23	1949
AMOM	−0.02 a	−2.60	0.08	1.01	0.10 a	2.73	0.12 b	2.08	0.14 a	3.50	0.08 b	2.24	0.19	1659
QRFT	−0.01 b	−2.43	0.07	1.17	0.07 b	2.15	0.17 a	4.42	0.10 b	2.56	0.11 a	2.85	0.19	1659
CHAT	0.02	1.24	0.10	1.38	0.13 b	2.49	0.09	1.13	0.05	1.34	0.05	1.02	0.16	653
AIPI	0.06 a	5.72	0.11 b	2.11	0.08	1.42	0.15 a	2.63	0.00	0.03	0.05	0.98	0.15	391
ALAI	0.03 b	2.48	0.06	0.79	0.25 a	2.88	0.07	1.16	0.05	0.91	0.1 b	2.05	0.24	432
FDTX	0.00	0.08	0.10	1.39	0.25 a	5.03	0.09 c	1.91	0.18 a	4.39	0.09	1.57	0.35	637
FBOT	−0.01	−0.78	0.15 a	2.72	0.10 b	2.16	0.15 a	3.04	0.09 c	1.95	0.04	0.81	0.22	637
ECML	0.01 c	1.95	0.09 b	2.33	0.14 a	3.51	0.08 c	1.91	0.08 c	1.83	0.16 a	3.81	0.21	653
AGIX	0.06 a	3.13	0.20 a	3.71	0.13 b	2.40	0.09	1.49	0.14 b	2.42	−0.04	−0.77	0.23	361
AIS	0.13 b	2.17	0.12	0.99	0.19 a	3.04	0.03	0.25	−0.07	−1.16	0.13	1.57	0.17	265
AIFD	−0.01	−1.49	0.19 a	3.95	0.09 c	1.88	0.13 b	2.57	0.19 a	3.62	0.11 b	2.13	0.36	411
GROZ	0.04 a	3.30	0.09	1.43	0.10	1.38	0.02	0.32	−0.04	−0.62	0.06	0.89	0.12	263
AIYY	0.07 a	2.64	0.00	0.00	0.03	0.80	0.05	1.42	0.08 b	2.09	0.06	0.83	0.11	520
LRNZ	0.01	1.52	0.07	0.92	0.21 a	3.04	0.17 a	2.61	0.15	1.47	0.16 c	1.99	0.37	1463
DRAI	0.03 a	5.49	−0.02	−0.37	0.10 c	1.78	0.00	0.00	−0.01	−0.26	0.13 b	2.43	0.12	367
GPT	0.02 c	1.93	−0.01	−0.06	0.28 a	3.08	0.08	0.76	0.19 c	1.87	0.06	0.68	0.28	318
NEWZ	0.10 a	7.35	−0.09 c	−1.88	−0.05	−0.95	−0.02	−0.43	0.05	0.95	−0.02	−0.31	0.11	427
XDAT	0.02	1.04	0.17 a	5.95	0.01	0.18	0.00	0.01	−0.01	−0.38	−0.09 a	−2.88	0.14	1242
LQAI	0.00	−1.14	0.10	1.21	0.05	1.27	0.07	1.36	0.05	1.18	0.09 b	2.33	0.14	534
Mean	0.03	1.67	0.08	1.40	0.11	2.06	0.08	1.54	0.07	1.33	0.06	1.07	0.19	1189
Min	−0.02	−3.49	−0.09	−1.88	−0.05	−0.95	−0.04	−0.56	−0.07	−1.16	−0.10	−2.88	0.11	263
Max	0.13	7.35	0.31	5.95	0.28	5.03	0.17	4.42	0.19	4.39	0.16	3.81	0.37	4911
Panel B: Passive ETFs
AIQ	0.03 a	3.63	0.12 a	2.85	0.13 a	4.14	0.02	0.65	0.05	1.60	0.15 a	4.35	0.19	1914
BOTZ	0.03 b	2.40	0.04	0.90	0.13 a	3.94	0.08 b	2.31	0.07 b	2.43	0.08 a	2.78	0.15	2334
ARTY	0.02 a	3.67	0.20 a	3.22	0.17 a	4.33	0.14 a	3.18	0.15 a	3.80	0.16 a	4.43	0.51	1883
ROBO	0.00	−0.09	0.08 c	1.92	0.19 a	6.13	0.11 a	3.18	0.09 a	3.30	0.11 a	4.38	0.23	3062
ROBT	0.04 a	4.11	0.09	1.14	0.13 a	3.92	0.12 a	3.99	0.18 a	5.37	0.17 a	4.29	0.31	1971
IGPT	0.00	−0.69	−0.07	−0.81	0.01	0.32	0.10 c	1.66	0.10 c	1.81	0.07 b	2.09	0.13	5158
DRIV	−0.02 c	−1.85	0.03	0.38	0.11 c	1.98	0.11 b	2.41	0.10 b	2.48	0.06	1.31	0.15	1934
AIEQ	−0.01 a	−2.73	0.07	1.64	0.14 a	4.17	0.06 c	1.87	0.03	0.97	0.12 a	3.51	0.16	2057
AIVC	0.00	−0.24	0.18 a	4.93	0.19 a	6.61	0.12 a	3.71	0.16 a	6.53	0.13 a	4.92	0.43	2464
WTAI	0.10 a	5.15	0.10 a	3.03	−0.05	−1.56	0.08 b	2.25	0.05	1.39	0.05	1.46	0.12	1014
THNQ	0.01 c	1.82	0.19 a	4.72	0.15 a	4.47	0.18 a	5.35	0.13 a	4.54	0.12 a	4.42	0.44	1414
BUZZ	−0.03 a	−5.35	−0.11	−0.72	0.02	0.39	0.08 b	2.29	0.07 b	2.01	0.05 c	1.74	0.12	1209
XAIX	0.06 b	2.46	0.05	0.50	0.30 a	4.04	0.03	0.24	0.08	1.20	0.16 b	2.15	0.27	350
RWEM	0.03	1.13	0.20 b	2.32	0.15 b	2.16	0.06	0.90	0.00	0.02	0.20 a	3.69	0.26	1009
RWLC	−0.04 a	−2.76	0.19 c	1.81	0.13 b	2.58	0.05	0.67	0.11	1.49	0.11 b	2.54	0.25	1009
HEAL	−0.04 a	−4.63	0.11 a	3.00	0.19 a	5.90	0.12 a	3.53	0.17 a	4.87	0.11 a	3.79	0.33	1358
WISE	0.08 a	4.88	0.07	1.52	0.14 a	3.15	0.11 b	2.29	0.10 b	2.07	0.11 b	2.41	0.20	512
IBOT	0.18 a	9.69	−0.01	−0.33	0.09 b	2.28	0.11 a	2.68	0.02	0.41	0.02	0.51	0.12	682
FAI	0.11 a	5.37	0.10 c	1.72	0.10	1.55	0.10	1.46	−0.03	−0.51	0.06	0.93	0.14	272
BOTT	0.30 a	3.98	−0.13	−1.23	0.13 c	1.95	0.02	0.19	0.05	0.67	0.07	1.13	0.14	421
KCAI	0.08	1.15	0.26	1.31	0.16	0.95	0.09	0.59	−0.06	−0.39	0.00	−0.04	0.23	332
BULD	−0.01	−0.26	0.38 a	2.64	0.19 b	2.59	−0.06	−0.85	−0.09	−1.62	−0.02	−0.32	0.31	913
Mean	0.04	1.40	0.10	1.66	0.13	3.00	0.08	2.02	0.07	2.02	0.10	2.57	0.24	1512
Min	−0.04	−5.35	−0.13	−1.23	−0.05	−1.56	−0.06	−0.85	−0.09	−1.62	−0.02	−0.32	0.12	272
Max	0.30	9.69	0.38	4.93	0.30	6.61	0.18	5.35	0.18	6.53	0.20	4.92	0.51	5158

This table presents the results of an autoregressive model via which the premium of each AI ETF is regressed on its lagged values up to five days. The estimation period of each AI ETF spans from its launch date up to 31 December 2025 (see Table 1 for launch dates). ^a indicates statistical significance at 1%; ^b indicates statistical significance at 5%; ^c indicates statistical significance at 10%.

. Regarding the one-lag premium of active AI ETFs, just seven significantly positive estimates and only one significantly negative coefficient are obtained. In the case of passive AI ETFs, 11 estimates are significantly positive. Moreover, most of the two-lag premium estimates are positive and significant (17 and 16 for active and passive AI ETFs, respectively), while no significantly negative estimate is produced. The influence of the three-lag premium is less strong, but still significantly positive for 11 active and 14 passive AI ETFs. Even weaker, but still positive, is the impact of the four-lag premium for 10 active ETFs and 11 passive ETFs. Finally, in nine cases, the five-lag premium is positive and significant for active ETFs and in one case significantly negative. For passive AI ETFs, the five-lag premium has 15 significantly positive estimates.

Overall, the results of Model (6) indicate that the premium of AI ETFs is quite persistent on a daily (and more than daily) basis. Given the positive sign for the majority of statistically significant estimates of the premium’s five-lag values considered in the regression model, the premium of an AI ETF on one day is likely to continue up to five days, at a minimum, based on our research approach. To our view, profitable investing strategies could be built on the evidence about premium persistence by executing efficient arbitrage techniques from the one trading day to the next.

Differences in premium persistence between active and passive AI ETFs can be interpreted through a combination of structural features and market frictions that limit the effectiveness of arbitrage. While, in theory, ETF arbitrage mechanisms should eliminate deviations from NAV, in practice several constraints prevent this adjustment from being instantaneous or costless, particularly in the context of thematic AI investing.

First, limits to arbitrage arise from incomplete transparency and informational frictions. Passive AI ETFs typically track rules-based indices and disclose their holdings on a daily basis, allowing authorized participants to observe the underlying basket with precision and implement low-risk arbitrage strategies. In contrast, active AI ETFs often provide only partial or lagged disclosure. This opacity introduces model uncertainty regarding the true composition and valuation of the portfolio, increasing the risk faced by arbitrageurs. As a result, arbitrage becomes less attractive, slower to implement, and conditional on wider mispricing thresholds, allowing premiums to persist for longer periods.

Second, liquidity constraints and replication frictions play a central role. Passive AI ETFs tend to hold large-cap, highly liquid technology stocks, which facilitates low-cost replication and efficient arbitrage. By contrast, active AI ETFs frequently invest in smaller, less liquid firms, including emerging AI innovators or niche technology companies. In such cases, arbitrageurs face higher transaction costs, price impact, and execution risk when trading the underlying securities. These frictions widen the no-arbitrage band, meaning that deviations from NAV must be sufficiently large to compensate for costs, thereby allowing smaller but persistent premiums to remain uncorrected.

Third, portfolio turnover and valuation uncertainty further weaken arbitrage efficiency. Active AI ETFs often rebalance dynamically to capture new opportunities in a rapidly evolving sector. Frequent changes in holdings complicate the accurate and timely estimation of NAV, especially when underlying securities are thinly traded or subject to asynchronous price updates. This creates stale pricing effects, whereby ETF prices, reflecting real-time trading, adjust more quickly than the reported NAV. Consequently, observed premiums may partly reflect informational lags rather than pure mispricing.

Fourth, investor sentiment and thematic demand are particularly important in the case of AI-related assets. Strong optimism about artificial intelligence, amplified by media coverage and technological breakthroughs, can generate sustained demand for AI ETFs. This demand may be price-insensitive in the short run, especially among retail investors or sentiment-driven flows, pushing ETF prices above NAV. Importantly, when demand surges rapidly, the supply response through the creation mechanism may lag, allowing premiums to persist. This mechanism is consistent with behavioral finance explanations in which attention-driven buying pressure leads to temporary but recurring deviations from fundamental values.

Fifth, information asymmetry and expectation effects contribute to the persistence of premiums. ETF prices incorporate forward-looking expectations about the growth potential of AI companies, whereas NAV is based on current or slightly lagged underlying prices. This divergence is more pronounced in active AI ETFs, where portfolio composition is less transparent. As a result, ETF prices may embed anticipatory valuations, effectively capitalizing expected future performance before it is reflected in the underlying assets.

Finally, the creation and redemption basket structure introduces additional frictions. While passive ETFs typically use pro rata baskets that closely mirror the underlying portfolio, active ETFs often employ custom baskets, allowing managers to include or exclude specific securities during the creation/redemption process. Although this provides flexibility, it also reduces the ability of arbitragers to perfectly replicate the underlying portfolio, increasing hedging error and transaction costs. This imperfect replication further weakens arbitrage effectiveness and contributes to the persistence of premiums.

Taken together, these mechanisms suggest that the observed premiums in both active and passive AI ETFs are not inconsistent with market efficiency per se, but rather reflect limits to arbitrage, liquidity constraints, and demand-driven pressures that are particularly pronounced in thematic and innovation-driven market segments.

6.3. Regression Analysis of Returns

The results of Model (7) on the impact of premium and intraday volatility on returns are presented in

Table 5. Return regression analysis.

Symbol	Const	t-Stat	Ret(-1)	t-Stat	Pr(-1)	t-Stat	Pr(-2)	t-Stat	IntVol	t-Stat	IntVol(-1)	t-Stat	R2	Obs
Panel A: Active ETFs
BAI	0.72	1.02	−0.06	−0.62	−1.05	−1.00	−0.81	−1.28	−0.45	−1.42	0.28 c	1.73	0.18	296
ARKQ	0.41 a	4.33	−0.03	−1.09	−0.17	−1.51	−0.87 a	−7.68	−0.39 a	−6.32	0.23 a	5.23	0.16	2828
IETC	0.38 a	2.68	−0.12 a	−2.90	−0.51 c	−1.70	−0.22	−0.78	−0.39 a	−3.91	0.21 a	3.57	0.17	1952
AIVL	0.07 c	1.78	−0.08 b	−2.30	−1.61 a	−4.02	−0.15	−0.45	−0.08 b	−2.25	0.02	0.78	0.14	4914
AIVI	0.04	1.12	0.05 b	2.17	1.17 a	29.41	−0.37 a	−6.92	−0.21 a	−4.25	0.17 a	3.76	0.50	4914
IEDI	0.17 b	2.09	−0.10 b	−2.02	0.26 b	2.33	−0.50 a	−3.52	−0.27 a	−2.79	0.11	1.43	0.14	1952
AMOM	0.35 a	3.61	−0.10 b	−2.58	0.35	1.30	−0.36	−0.97	−0.40 a	−3.26	0.13 c	1.77	0.17	1662
QRFT	0.29 a	6.90	−0.13 a	−5.22	−0.32 c	−1.88	−0.18	−1.05	−0.47 a	−12.28	0.13 a	3.23	0.20	1662
CHAT	0.07	0.26	0.07	1.37	3.82 a	10.21	−1.64 a	−6.09	−0.23 c	−1.81	0.23 b	2.38	0.57	656
AIPI	0.25	0.94	−0.08	−1.00	−0.98	−1.11	−0.87	−1.43	−0.52 a	−2.87	0.42 a	3.21	0.22	394
ALAI	0.32	0.96	−0.04	−0.49	2.04 b	2.11	−1.87 c	−1.81	−0.35	−1.48	0.22 b	2.03	0.19	435
FDTX	0.16	0.36	0.03	0.36	3.55 a	9.54	−2.23 a	−5.16	−0.19	−0.71	0.15	0.94	0.30	640
FBOT	−0.18	−0.68	0.11	1.46	2.19 a	12.77	−0.72 a	−3.75	0.12	0.51	0.11	0.82	0.59	640
ECML	0.04	0.44	0.01	0.18	−0.47	−0.28	−0.03	−0.03	0.14	0.66	−0.08	−0.67	0.11	656
AGIX	0.58	1.64	0.00	0.02	1.32 a	4.59	−1.68 a	−3.39	−0.51 a	−3.06	0.26	1.62	0.30	364
AIS	−0.03	−0.19	0.05	0.83	3.81 a	18.02	−1.86 a	−6.60	−0.14 b	−2.13	0.03	0.47	0.67	268
AIFD	0.32	0.78	−0.08	−1.04	−0.37	−0.43	−0.99	−1.27	−0.30	−1.27	0.15	1.30	0.15	414
GROZ	0.09	0.26	−0.15	−1.55	0.90	1.63	−1.29 a	−2.84	−0.12	−0.45	0.13	1.03	0.14	266
AIYY	0.98 c	1.69	−0.07	−1.43	−0.51	−1.10	0.15	0.52	−0.48 a	−2.72	0.10	0.91	0.19	523
LRNZ	0.48 a	3.01	0.00	0.12	−0.47	−0.99	0.64 c	1.67	−0.25 a	−3.38	0.06	1.08	0.13	1466
DRAI	0.22 b	2.20	−0.08	−1.36	−0.57	−0.44	−0.48	−0.62	−0.24	−1.62	0.03	0.43	0.14	370
GPT	0.17	0.97	−0.05	−0.58	−0.16	−0.52	−0.15	−0.39	−0.02	−0.08	−0.07	−0.61	0.11	321
NEWZ	0.13	0.94	−0.04	−0.62	0.97 b	2.24	−0.79 a	−2.79	−0.33	−1.41	0.10	0.60	0.16	430
XDAT	0.13	1.32	0.00	−0.01	0.14	1.52	−0.95 a	−5.32	−0.06	−0.54	−0.06	−0.80	0.16	1245
LQAI	0.05	0.47	−0.10	−1.07	−0.34	−0.63	−1.09 b	−2.21	−0.23	−1.59	0.34	1.36	0.15	537
Mean	0.25	1.55	−0.04	−0.77	0.52	3.20	−0.77	−2.57	−0.25	−2.42	0.14	1.50	0.24	1192
Min	−0.18	−0.68	−0.15	−5.22	−1.61	−4.02	−2.23	−7.68	−0.52	−12.28	−0.08	−0.80	0.11	266
Max	0.98	6.90	0.11	2.17	3.82	29.41	0.64	1.67	0.14	0.66	0.42	5.23	0.67	4914
Panel B: Passive ETFs
AIQ	0.14	0.85	−0.04	−0.96	2.49 a	9.45	−0.95 a	−4.18	−0.36 a	−3.81	0.26 a	4.40	0.35	1917
BOTZ	−0.06	−0.94	0.10 a	3.49	2.09 a	30.18	−0.65 a	−8.51	−0.14 a	−3.14	0.16 a	3.81	0.67	2337
ARTY	0.21	0.98	0.00	−0.06	2.31 a	9.39	−1.73 a	−7.19	−0.31 a	−2.86	0.17 b	2.38	0.23	1886
ROBO	0.01	0.17	0.19 a	7.68	2.32 a	36.28	−1.13 a	−14.81	−0.07 b	−2.57	0.09 a	3.49	0.61	3065
ROBT	0.30	1.59	−0.05	−1.19	1.06 a	3.88	−0.98 a	−4.73	−0.39 a	−3.49	0.21 a	2.83	0.18	1974
IGPT	0.24 a	4.30	−0.06 a	−2.69	−0.32 b	−2.27	−0.75 a	−5.50	−0.13 a	−3.30	0.01	0.46	0.14	5162
DRIV	0.23	1.64	0.00	−0.01	2.32 a	5.79	−0.68 b	−2.56	−0.35 a	−4.16	0.27 a	3.48	0.39	1937
AIEQ	0.26 c	1.89	−0.07	−1.37	−1.07 a	−3.91	−0.13	−0.49	−0.32 a	−2.85	0.15 b	2.08	0.15	2060
AIVC	0.18 b	2.58	0.12 a	3.92	2.20 b	15.47	−1.65 a	−13.16	−0.11 b	−2.50	0.02	0.61	0.34	2467
WTAI	0.00	−0.01	0.08 b	2.02	1.22 a	6.14	−0.93 a	−7.07	−0.11	−0.77	0.11	1.30	0.25	1017
THNQ	0.38 c	1.82	0.06 c	1.72	2.96 a	7.97	−2.05 a	−7.29	−0.32 a	−3.19	0.12 c	1.75	0.25	1417
BUZZ	0.32	1.24	0.00	0.07	0.74 c	1.91	−0.78 a	−2.92	−0.26 a	−2.66	0.14 c	1.94	0.13	1212
XAIX	0.04	0.18	0.16 b	2.59	4.95 a	6.34	−3.55 a	−6.25	−0.20 c	−1.72	0.10	1.42	0.53	353
RWEM	0.01	0.34	0.12 a	3.34	0.83 a	13.17	−0.69 a	−12.45	−0.04	−1.29	0.04	1.14	0.56	1012
RWLC	0.08	1.56	0.07 c	1.83	0.93 a	7.52	−0.89 a	−9.77	−0.11	−1.11	0.06	1.44	0.30	1012
HEAL	0.01	0.06	0.06 b	1.96	1.93 a	13.46	−1.07 a	−7.60	0.01	0.17	0.03	0.66	0.26	1361
WISE	−0.16	−0.58	0.03	0.60	4.33 a	8.47	−1.45 a	−2.96	−0.18	−1.44	0.08	0.92	0.34	515
IBOT	−0.25	−1.19	−0.06	−0.91	0.94	1.64	0.65	1.61	−0.17	−0.67	0.12	1.01	0.13	685
FAI	0.12	0.32	−0.10	−0.95	2.34 b	2.17	−2.30 a	−2.54	−0.12	−0.53	0.13	0.99	0.18	275
BOTT	−0.30 b	−2.34	−0.02	−0.30	2.04 a	12.32	−0.04	−0.19	−0.20 a	−2.05	−0.06	−0.67	0.55	424
KCAI	0.12	1.07	0.02	0.37	0.94 a	9.12	−0.68 a	−6.25	−0.07	−0.39	−0.22	−1.23	0.32	335
BULD	0.10 c	1.68	0.07 c	1.89	0.80 a	8.96	−0.90 a	−11.65	−0.20 c	−1.77	0.02	0.23	0.32	916
Mean	0.09	0.78	0.03	1.05	1.74	9.25	−1.06	−6.20	−0.19	−2.10	0.09	1.57	0.33	1512
Min	−0.30	−2.34	−0.10	−2.69	−1.07	−3.91	−3.55	−14.81	−0.39	−4.16	−0.22	−1.23	0.13	272
Max	0.38	4.30	0.19	7.68	4.95	36.28	0.65	1.61	0.01	0.17	0.27	4.40	0.67	5158

This table presents the results of a regression model via which the returns each AI ETF is regressed on its one-lag value, the concurrent and the one-lag premium, and the concurrent and the one-lagged intraday volatility. The estimation period of each AI ETF spans from its launch date up to 31 December 2025 (see Table 1 for launch dates). ^a indicates statistical significance at 1%; ^b indicates statistical significance at 5%; ^c indicates statistical significance at 10%.

. First, with respect to the relationship between current and lagged returns, five significantly negative and one positive and significant estimates are obtained for active ETFs. In the case of passive AI ETFs, 10 coefficients are significantly positive and 1 is negative and significant. These results indicate that there is not a monotonic relationship between current and lagged returns for all AI ETFs.

When it comes to concurrent premium, this factor has eight and two significantly positive and negative estimates, respectively, in the case of active ETFs. In the case of passive ETFs, 19 (2) positive (negative) and significant coefficients are obtained. These results reveal a significant relationship between current return and premium, especially for the passive AI ETFs. On the other hand, the relationship between current return and lagged premium is rather negative. In the case of active ETFs, 13 estimates are significantly negative (and only 1 is significantly positive), while the passive ETFs have 19 significantly negative coefficients and none that are significantly positive.

A negative impact of the current intraday volatility on current return is revealed too. Thirteen and fourteen coefficients are significantly negative for active and passive ETFs, respectively, while no positive and significant estimate is produced by the model. The opposite relationship applies to the lagged intraday volatility, at least for nine active and nine passive ETFs.

Overall, the results of Model (7) indicate that factors such as the current and lagged premium, the current and lagged intraday volatility, and, to some extent, lagged returns, can be quite important when trying to detect factors that can affect the return of AI ETFs. We deem that the positive relationship (on average) between return and premium, the negative relationship between return and lagged premium, and the opposite patterns found between return and intraday volatility could be the basis for prosperous investing strategies.

The negative contemporaneous and positive lagged relationship between volatility and returns can be interpreted through the lens of the volatility feedback hypothesis (Campbell & Hentschel, 1992). This framework posits that an increase in expected volatility raises the required risk premium, leading to an immediate decline in asset prices as investors demand higher compensation for bearing risk. Such dynamics provide a natural explanation for the observed negative contemporaneous relationship between volatility and returns.

Over time, however, as markets gradually incorporate information and risk premia adjust, part of the initial price decline may reverse, giving rise to a positive lagged association between volatility and subsequent returns. This intertemporal adjustment mechanism is consistent with the idea that volatility shocks can have both immediate pricing effects and delayed return implications.

Market microstructure effects may further reinforce this pattern. Short-term liquidity shocks, order-flow imbalances, bid–ask bounce, and other trading frictions can temporarily depress prices, particularly in ETF markets where trading activity and investor flows may induce deviations from underlying asset values. As these frictions dissipate and liquidity conditions normalize, prices may partially recover, contributing to the observed positive lagged return dynamics.

Finally, given that volatility and returns are jointly determined in equilibrium, the documented relationship may also reflect endogenous feedback effects between risk perceptions and AI ETF prices. This highlights the importance of incorporating lagged dynamics when analyzing ETF returns, as doing so helps disentangle short-run price pressure from adjustments in risk premia.

6.4. Performance Regression Analysis

The results of Model (8) are discussed in this section. The model has been applied both with trade price and NAV returns. The estimates of the first version of the model are provided in

Table 6. Performance regression analysis (price returns).

Symbol	Const	t-Stat	Mkt	t-Stat	SMB	t-Stat	HML	t-Stat	CMA	t-Stat	RMW	t-Stat	MOM	t-Stat	R2	Obs
Panel A: Active ETFs
BAI	0.06	1.14	1.26 a	26.97	0.18 b	2.08	−0.77 a	−10.27	0.08	0.74	−0.32 a	−3.20	0.44 a	7.36	0.87	297
ARKQ	0.03 c	1.80	1.11 a	44.81	0.52 a	15.23	−0.32 a	−9.60	−0.23 a	−4.16	−0.53 a	−13.28	−0.01	−0.72	0.77	2829
IETC	0.02 b	2.17	1.14 a	133.71	−0.10 a	−6.22	−0.36 a	−24.40	−0.17 a	−7.01	0.00	−0.24	0.05 a	5.62	0.92	1953
AIVL	−0.01 b	−2.30	0.86 a	89.87	0.01	0.53	0.19 a	13.19	0.21 a	9.16	0.13 a	7.96	−0.15 a	−17.44	0.91	4915
AIVI	−0.03 b	−2.54	0.91 a	96.02	−0.09 a	−4.96	0.16 a	9.39	0.14 a	4.50	−0.02	−0.90	−0.05 a	−4.63	0.71	4915
IEDI	0.00	−0.01	0.86 a	54.51	0.16 a	6.50	−0.15 a	−6.32	−0.03	−0.74	0.20 a	7.02	−0.03 c	−1.84	0.81	1953
AMOM	−0.01	−0.42	1.04 a	43.60	0.11 a	3.09	−0.19 a	−5.15	−0.03	−0.54	−0.18 a	−3.01	0.35 a	18.46	0.77	1663
QRFT	0.01	0.79	0.93 a	76.17	0.04 b	2.24	−0.16 a	−8.17	−0.01	−0.44	−0.04	−1.58	0.11 a	9.80	0.92	1663
CHAT	0.03	1.04	1.41 a	32.18	−0.01	−0.19	−0.54 a	−6.73	−0.15	−1.12	−0.24 a	−2.73	0.20 a	2.73	0.81	657
AIPI	−0.14 a	−3.38	0.95 a	22.65	−0.07	−1.03	−0.26 a	−4.05	0.16 c	1.89	−0.37 a	−4.11	0.21 a	4.06	0.74	395
ALAI	0.06 b	2.18	1.31 a	41.50	0.01	0.18	−0.60 a	−12.38	0.11 c	1.70	0.00	−0.06	0.44 a	11.37	0.90	436
FDTX	−0.02	−0.80	1.25 a	34.18	−0.06	−1.21	−0.34 a	−7.04	−0.17 b	−2.26	−0.42 a	−6.66	0.18 a	5.10	0.86	641
FBOT	−0.02	−0.97	1.03 a	36.60	0.19 a	4.26	−0.29 a	−6.63	0.06	1.02	−0.12	−2.13	0.05	1.44	0.78	641
ECML	0.00	0.18	0.93 a	45.33	0.77 a	24.05	0.26 a	8.30	0.07 c	1.77	0.49 a	12.30	−0.05 b	−2.07	0.87	657
AGIX	0.03	0.85	1.18 a	31.71	0.15 b	2.34	−0.64 a	−10.91	−0.06	−0.76	−0.36 a	−4.47	0.22 a	4.72	0.88	365
AIS	0.12 c	1.93	1.27 a	19.89	0.25 c	1.82	−0.70 a	−4.54	0.32	1.19	−0.33 c	−1.89	0.23	1.64	0.79	269
AIFD	0.01	0.32	1.33 a	30.71	0.07	1.08	−0.59 a	−8.91	0.05	0.48	−0.12	−1.31	0.33 a	5.28	0.89	415
GROZ	0.03	1.49	1.12 a	57.43	−0.03	−0.75	−0.37 a	−10.78	0.05	1.03	0.04	0.95	0.20 a	7.82	0.96	267
AIYY	−0.49 a	−4.03	1.17 a	8.54	0.91 a	4.21	−0.52 b	−2.49	0.48 c	1.70	−1.10 a	−3.95	−0.08	−0.49	0.35	524
LRNZ	0.03	1.28	1.03 a	49.77	0.22 a	5.38	−0.64 a	−18.18	−0.65 a	−11.28	−1.03 a	−20.60	−0.04 c	−1.87	0.83	1467
DRAI	0.02	0.57	0.60 a	6.21	0.25 b	2.57	−0.37 a	−4.67	0.09	1.04	0.04	0.43	0.04	0.48	0.58	371
GPT	0.01	0.15	0.88 a	13.92	−0.03	−0.25	−0.18 c	−1.91	0.19	1.50	−0.24 b	−2.26	0.01	0.11	0.75	322
NEWZ	−0.03	−1.18	0.74 a	15.48	0.17 b	2.61	0.06	1.18	0.07	1.23	−0.20 a	−3.29	0.15 a	3.89	0.75	431
XDAT	0.00	−0.12	1.01 a	32.33	−0.01	−0.32	−0.42 a	−11.33	−0.51 a	−8.70	−0.61 a	−9.91	0.04	1.48	0.78	1246
LQAI	0.00	−0.36	0.99 a	63.63	−0.06 b	−2.62	−0.02	−0.75	−0.04	−1.14	−0.12 a	−3.72	0.08 a	4.35	0.92	538
Mean	−0.01	−0.01	1.05	44.31	0.14	2.42	−0.31	−5.73	0.00	−0.37	−0.22	−2.43	0.12	2.66	0.80	1193
Min	−0.49	−4.03	0.60	6.21	−0.10	−6.22	−0.77	−24.40	−0.65	−11.28	−1.10	−20.60	−0.15	−17.44	0.35	267
Max	0.12	2.18	1.41	133.71	0.91	24.05	0.26	13.19	0.48	9.16	0.49	12.30	0.44	18.46	0.96	4915
Panel B: Passive ETFs
AIQ	0.02	1.57	1.04 a	90.99	0.05 b	2.26	−0.33 a	−16.84	−0.26 a	−7.84	−0.21 a	−7.49	0.00	0.03	0.86	1918
BOTZ	−0.01	−0.45	1.12 a	53.54	0.18 a	5.62	−0.20 a	−6.78	−0.11 b	−2.23	−0.22 a	−5.88	−0.04 c	−1.95	0.79	2338
ARTY	0.00	0.23	1.05 a	43.52	0.28 a	9.32	−0.33 a	−8.82	−0.08	−1.14	−0.40 a	−10.34	−0.02	−1.34	0.83	1887
ROBO	0.00	−0.37	1.03 a	74.09	0.30 a	15.08	−0.16 a	−8.49	0.02	0.70	−0.14 a	−4.98	−0.06 a	−4.91	0.83	3066
ROBT	−0.01	−0.61	1.04 a	92.93	0.36 a	16.31	−0.24 a	−12.30	−0.06 b	−2.00	−0.28 a	−10.48	−0.04 a	−3.31	0.87	1975
IGPT	0.02 b	2.34	0.96 a	51.70	0.29 a	11.41	−0.42 a	−18.17	−0.26 a	−6.77	−0.47 a	−15.10	0.12 a	8.64	0.79	5163
DRIV	0.00	−0.09	1.10 a	33.50	0.32 a	7.33	−0.13 a	−3.17	0.06	0.98	−0.17 a	−4.40	−0.13 a	−4.66	0.78	1938
AIEQ	−0.01	−0.72	0.99 a	55.33	0.23 a	7.41	−0.19 a	−7.84	−0.01	−0.13	−0.31 a	−10.90	0.05 b	2.29	0.85	2061
AIVC	0.00	0.26	0.98 a	36.41	0.17 a	5.17	−0.34 a	−8.88	−0.18 a	−2.82	−0.67 a	−17.46	−0.03	−1.43	0.73	2468
WTAI	0.00	−0.14	1.17 a	29.09	0.25 a	4.35	−0.28 a	−5.06	−0.15 c	−1.92	−0.57 a	−8.04	0.19 a	5.25	0.76	1018
THNQ	0.03 b	2.05	1.14 a	72.77	0.11 a	4.31	−0.39 a	−17.14	−0.30 a	−8.58	−0.58 a	−18.71	−0.02	−1.13	0.89	1418
BUZZ	0.02	0.68	1.28 a	54.50	0.24 a	6.17	−0.18 a	−5.06	−0.33 a	−6.54	−0.81 a	−18.19	−0.02	−0.91	0.85	1213
XAIX	0.02	0.92	1.10 a	50.98	−0.07 c	−1.85	−0.23 a	−6.73	−0.06	−1.36	−0.22 a	−4.68	0.09 a	3.25	0.93	354
RWEM	−0.01	−0.46	0.56 a	20.64	−0.02	−0.50	0.02	0.50	0.07	1.24	−0.12 b	−2.32	0.03	1.02	0.36	1013
RWLC	0.01	0.33	0.76 a	42.20	0.00	0.04	−0.03	−1.09	0.08 b	2.07	−0.03	−0.73	0.13 a	6.77	0.69	1013
HEAL	−0.04	−1.58	0.78 a	30.08	0.54 a	12.64	−0.43 a	−10.78	0.02	0.27	−0.57 a	−11.31	−0.08 a	−3.15	0.71	1362
WISE	0.01	0.15	1.30 a	25.47	0.36 a	4.52	−0.54 a	−7.00	0.11	1.01	−0.84 a	−8.11	0.15 b	2.41	0.77	516
IBOT	−0.01	−0.57	1.16 a	39.54	0.19 a	4.13	−0.19 a	−4.27	0.05	0.84	0.00	0.08	0.05	1.41	0.77	686
FAI	0.06 c	1.73	1.25 a	34.93	0.06	0.95	−0.71 a	−11.29	0.00	−0.03	−0.22 a	−2.97	0.18 a	4.01	0.91	276
BOTT	0.02	0.42	1.17 a	21.22	0.26 a	2.80	−0.29 a	−3.45	0.22 c	1.94	−0.34 a	−2.95	0.18 b	2.64	0.69	425
KCAI	0.13	1.56	0.21 b	2.48	−0.04	−0.26	−0.10	−0.74	0.14	0.75	−0.18	−1.01	−0.27 b	−2.55	0.17	336
BULD	−0.01	−0.23	0.99 a	24.76	0.46 a	6.38	−0.16 b	−2.53	−0.03	−0.31	−0.17 b	−2.11	0.12 a	2.82	0.55	917
Mean	0.01	0.32	1.01	44.58	0.21	5.62	−0.27	−7.54	−0.05	−1.45	−0.34	−7.64	0.03	0.69	0.74	1512
Min	−0.04	−1.58	0.21	2.48	−0.07	−1.85	−0.71	−18.17	−0.33	−8.58	−0.84	−18.71	−0.27	−4.91	0.17	272
Max	0.13	2.34	1.30	92.93	0.54	16.31	0.02	0.50	0.22	2.07	0.00	0.08	0.19	8.64	0.93	5158

This table presents the results of a six-factor performance regression model via which the daily excess return of AI ETFs (calculated in close trade prices terms) is regressed on the excess return of the Russell 3000 Index, and the Fama and French (2015) SMB (Small Minus Big) factor, HML (High Minus Low book-to-price ratio) factor, CMA (Conservative Minus Aggressive) factor, RMW (Robust Minus Weak) factor, and Carhart’s MOM (momentum) factor. The estimation period of each AI ETF spans from its launch date up to 31 December 2025 (see Table 1 for launch dates). ^a indicates statistical significance at 1%; ^b indicates statistical significance at 5%; ^c indicates statistical significance at 10%.

. We begin our analysis by noting that the majority of the AI ETFs examined cannot achieve material alphas against the selected market index. In the case of active ETFs, only four significantly positive alphas are provided by the regression model and four that are significantly negative. In the case of passive ETFs, only three funds achieve positive and significant alphas, while none presents significantly negative alphas. Overall, these results indicate that, on average, the AI ETFs rather go with the flow, without beating the market or being beaten by the market. Our evidence challenges the findings of Malhotra (2025), who reports that the AI ETFs can provide their investors with positive and significant alphas.

On systematic risk, an average beta of 1.05 is estimated for active AI ETFs, while the respective average coefficient for passive ETFs is equal to 1.01. All individual betas are statistically significant and positive. Moreover, 15 active ETFs have betas that are higher than unity, implying that the respective ETFs move more aggressively than the market. Fourteen passive ETFs do so too. Aggressiveness helps funds realize more gains during bull markets, but punishes them with greater losses during bear markets.

Regarding the size factor, 13 active AI ETFs present significantly positive estimates and only 3 present negative and significant SMB estimates. In the case of passive ETFs, 17 out of 22 SMB estimates are positive and significant, and only 1 is significantly negative. These results establish a positive relationship between the performance of AI ETFs and the size factor suggested by Fama and French (1993) for the stock market in the U.S.

On the value factor suggested by Fama and French (1993), the results indicate a relevant negative impact on AI ETFs’ performance. In particular, 20 HML estimates are negative and significant in the case of active ETFs and only 3 are positive and significant. The corresponding numbers for passive AI ETFs are 19 and 0.

As far as the conservativeness factor of Fama and French (2015) is concerned, the results in Table 6 reveal a rather weak and unsystematic influence on AI ETFs’ return. The average CMA coefficient of active ETFs is equal to zero, while the respective average term of passive ETFs is equal to −0.05. At the fund level, six active AI ETFs present significantly positive estimates and five funds have significantly negative CMA estimates. The respective numbers for passive ETFs are eight and two.

Contrary to conservativeness, the robustness factor offers quite significant estimates, both for active and passive ETFs. More specifically, 15 coefficients are significantly negative in the case of active AI ETFs and only 3 are positive and significant. In the case of passive ETFs, 19 estimates are negative and significant and none is significantly positive. Overall, these results accentuate a clear negative impact of the robustness factor on the performance of AI ETFs in the U.S.

Finally, the impact of the momentum factor included in Model (8) is positive for the majority of the AI ETFs examined, (i.e., for 15 active ETFs and 9 passive ETFs), while 11 ETFs present significantly negative MOM coefficients (5 active and 6 passive).

To summarize our results, we note that the regression model applied can sufficiently explain the performance of AI ETFs traded in the U.S. stock market. Most of the variables included in the model produce statistically significant estimates, either negative or positive, but at the average level, performance is found to be related to the size and momentum factors in a positive fashion, while the influence of the value and robustness factors on performance is negative. Finally, the contribution of the conservativeness factor to AI ETFs’ performance is less significant.

Before concluding this section, we should remind readers that the analysis above concerns the results of Model (8) obtained when the return of AI ETFs calculated in trade price terms is the models’ dependent variable.

Table 7. Performance regression analysis (NAVs).

Symbol	Const	t-Stat	Mkt	t-Stat	SMB	t-Stat	HML	t-Stat	CMA	t-Stat	RMW	t-Stat	MOM	t-Stat	R2	Obs
Panel A: Active ETFs
BAI	0.05	1.14	1.30 a	28.75	0.12	1.39	−0.77 a	−10.52	0.05	0.44	−0.30 a	−3.04	0.46 a	7.97	0.88	297
ARKQ	0.03 c	1.84	1.16 a	55.23	0.53 a	17.67	−0.31 a	−10.42	−0.24 a	−4.81	−0.52 a	−14.00	−0.01	−0.50	0.81	2829
IETC	0.02 b	2.16	1.16 a	81.22	−0.10 a	−4.91	−0.35 a	−17.54	−0.20 a	−5.48	0.01	0.29	0.06 a	5.76	0.94	1953
AIVL	−0.01 b	−2.50	0.88 a	92.15	0.00	−0.17	0.21 a	14.26	0.21 a	9.09	0.13 a	7.91	−0.15 a	−18.80	0.92	4915
AIVI	−0.01	−0.80	0.45 a	15.42	−0.08 c	−1.65	0.13 a	2.79	0.04	0.56	−0.09 c	−1.78	−0.10 a	−4.07	0.27	4915
IEDI	0.00	−0.28	0.90 a	67.84	0.16 a	7.76	−0.16 a	−7.27	−0.05	−1.28	0.24 a	9.57	−0.02	−1.51	0.87	1953
AMOM	−0.01	−0.55	1.08 a	72.30	0.05 c	1.81	−0.19 a	−7.31	−0.01	−0.32	−0.18 a	−4.80	0.34 a	20.87	0.80	1663
QRFT	0.01	0.71	0.97 a	157.90	0.01	0.87	−0.14 a	−12.92	−0.03	−1.61	−0.03 b	−2.02	0.11 a	15.96	0.95	1663
CHAT	0.04	1.25	1.22 a	23.26	−0.03	−0.39	−0.48 a	−5.98	−0.23 c	−1.69	−0.18 c	−1.97	0.19 b	2.59	0.77	657
AIPI	−0.14 a	−3.43	0.96 a	23.24	−0.09	−1.27	−0.28 a	−4.34	0.17 c	1.99	−0.38 a	−4.29	0.19 a	3.76	0.75	395
ALAI	0.06 b	2.38	1.27 a	43.81	0.01	0.28	−0.60 a	−13.59	0.10 c	1.75	−0.03	−0.56	0.44 a	12.46	0.91	436
FDTX	−0.02	−0.75	1.20 a	34.59	−0.07	−1.43	−0.33 a	−7.11	−0.23 a	−3.47	−0.42 a	−6.92	0.14 a	4.17	0.86	641
FBOT	−0.02	−0.68	0.82 a	24.50	0.08	1.52	−0.20 a	−4.15	−0.08	−1.11	−0.07	−1.05	−0.02	−0.39	0.64	641
ECML	0.00	0.16	0.93 a	45.45	0.77 a	24.03	0.27 a	8.58	0.07	1.60	0.49 a	12.27	−0.05 c	−1.91	0.87	657
AGIX	0.03	0.89	1.17 a	35.10	0.06	1.08	−0.52 a	−9.91	0.00	0.07	−0.36 a	−5.04	0.20 a	4.76	0.89	365
AIS	0.12 c	1.99	0.99 a	10.99	0.23 c	1.71	−0.63 a	−4.17	0.32	1.26	−0.33 c	−1.92	0.24 c	1.77	0.73	269
AIFD	0.01	0.37	1.34 a	29.38	0.07	1.03	−0.61 a	−8.84	0.06	0.66	−0.10	−1.21	0.32 a	5.29	0.89	415
GROZ	0.03 c	1.70	1.12 a	66.67	−0.03	−0.93	−0.36 a	−12.08	0.02	0.48	0.04	1.04	0.20 a	9.02	0.97	267
AIYY	−0.49 a	−3.85	1.19 a	8.33	0.94 a	4.19	−0.59 a	−2.72	0.45	1.54	−1.23 a	−4.27	−0.11	−0.63	0.36	524
LRNZ	0.03	1.20	1.10 a	46.32	0.17 a	3.99	−0.61 a	−16.72	−0.65 a	−9.73	−1.03 a	−19.51	−0.04 c	−1.65	0.84	1467
DRAI	0.02	0.58	0.60 a	6.11	0.24 b	2.42	−0.37 a	−4.53	0.09	0.98	0.02	0.28	0.02	0.29	0.59	371
GPT	0.01	0.19	0.88 a	13.76	−0.03	−0.21	−0.19 b	−2.00	0.26 b	2.08	−0.26 b	−2.49	0.01	0.19	0.75	322
NEWZ	−0.03	−1.31	0.75 a	15.47	0.18 a	2.76	0.06	1.10	0.08	1.56	−0.17 a	−3.04	0.16 a	4.41	0.79	431
XDAT	−0.01	−0.26	1.08 a	52.76	−0.02	−0.65	−0.41 a	−13.73	−0.54 a	−12.34	−0.60 a	−15.63	0.03	1.34	0.86	1246
LQAI	−0.01	−0.42	1.00 a	68.84	−0.07 a	−2.95	−0.01	−0.56	−0.03	−1.18	−0.12 a	−4.02	0.08 a	4.75	0.93	538
Mean	−0.01	0.07	1.02	44.78	0.12	2.32	−0.30	−5.99	−0.02	−0.76	−0.22	−2.65	0.11	3.04	0.79	1193
Min	−0.49	−3.85	0.45	6.11	−0.10	−4.91	−0.77	−17.54	−0.65	−12.34	−1.23	−19.51	−0.15	−18.80	0.27	267
Max	0.12	2.38	1.34	157.90	0.94	24.03	0.27	14.26	0.45	9.09	0.49	12.27	0.46	20.87	0.97	4915
Panel B: Passive ETFs
AIQ	0.02 c	1.81	1.00 a	104.38	0.01	0.55	−0.26 a	−16.15	−0.29 a	−10.66	−0.17 a	−7.22	0.01	0.90	0.89	1918
BOTZ	0.01	0.43	0.70 a	42.14	0.14 a	4.59	−0.15 a	−5.24	−0.14 a	−3.07	−0.26 a	−6.77	−0.04 b	−2.05	0.53	2338
ARTY	0.01	0.33	0.99 a	36.30	0.25 a	7.64	−0.33 a	−8.52	−0.08	−1.07	−0.41 a	−10.87	−0.04 b	−2.19	0.82	1887
ROBO	0.01	0.52	0.72 a	34.80	0.27 a	10.33	−0.12 a	−4.72	−0.03	−0.69	−0.15 a	−4.57	−0.08 a	−4.84	0.65	3066
ROBT	−0.01	−0.74	1.04 a	67.81	0.33 a	12.28	−0.25 a	−11.23	−0.05	−1.14	−0.26 a	−9.02	−0.05 a	−4.00	0.89	1975
IGPT	0.02 b	2.41	1.02 a	86.55	0.29 a	17.40	−0.42 a	−21.75	−0.25 a	−7.24	−0.43 a	−17.85	0.12 a	9.55	0.85	5163
DRIV	0.00	−0.01	1.01 a	44.79	0.32 a	10.18	−0.06 c	−1.71	−0.05	−0.89	−0.15 a	−3.72	−0.09 a	−4.61	0.78	1938
AIEQ	−0.01	−0.88	1.03 a	62.42	0.26 a	10.35	−0.19 a	−8.71	−0.02	−0.43	−0.28 a	−10.71	0.06 a	2.82	0.87	2061
AIVC	0.01	0.41	0.87 a	35.10	0.18 a	5.33	−0.33 a	−9.70	−0.20 a	−3.46	−0.62 a	−17.23	−0.03	−1.32	0.72	2468
WTAI	−0.01	−0.23	1.13 a	37.69	0.20 a	4.08	−0.27 a	−5.84	−0.20 a	−3.06	−0.48 a	−8.35	0.12 a	3.98	0.81	1018
THNQ	0.03 b	2.24	1.10 a	76.00	0.10 a	4.21	−0.37 a	−17.60	−0.31 a	−9.43	−0.57 a	−19.88	−0.02	−1.26	0.90	1418
BUZZ	0.02	0.69	1.29 a	57.15	0.24 a	6.47	−0.19 a	−5.49	−0.34 a	−6.91	−0.79 a	−18.60	−0.02	−0.85	0.86	1213
XAIX	0.02	1.08	1.00 a	46.13	−0.07 c	−1.88	−0.19 a	−5.62	−0.08 c	−1.78	−0.22 a	−4.76	0.10 a	3.66	0.92	354
RWEM	−0.01	−0.20	0.21 a	7.18	−0.07	−1.32	0.11 b	2.47	−0.11 c	−1.76	−0.15 a	−2.71	0.01	0.30	0.18	1013
RWLC	0.01	0.65	0.77 a	78.94	−0.02	−0.94	0.00	−0.07	0.02	1.09	−0.03 c	−1.73	0.13 a	12.46	0.89	1013
HEAL	−0.04 c	−1.75	0.74 a	29.80	0.53 a	13.37	−0.40 a	−10.90	−0.02	−0.28	−0.54 a	−11.85	−0.08 a	−3.05	0.73	1362
WISE	0.01	0.25	1.22 a	25.01	0.41 a	5.25	−0.52 a	−7.03	0.09	0.88	−0.77 a	−7.77	0.17 a	2.87	0.77	516
IBOT	−0.02	−0.62	1.18 a	41.78	0.17 a	3.92	−0.20 a	−4.66	0.08	1.41	0.01	0.26	0.05	1.46	0.79	686
FAI	0.06 c	1.72	1.25 a	24.19	−0.01	−0.06	−0.65 a	−8.61	0.00	−0.04	−0.24 a	−2.72	0.16 a	3.19	0.92	276
BOTT	0.04	0.66	0.76 a	12.33	0.31 a	3.08	−0.18 c	−1.94	0.08	0.64	−0.26 b	−2.04	0.25 a	3.34	0.46	425
KCAI	0.14 b	2.10	0.27 c	1.79	0.07	0.61	−0.21 c	−1.95	0.20	1.34	0.05	0.34	−0.08	−0.91	0.13	336
BULD	−0.01	−0.48	0.98 a	24.44	0.38	7.40	−0.19 a	−3.59	−0.13 c	−1.73	−0.15 b	−2.55	−0.03	−0.75	0.72	917
Mean	0.01	0.47	0.92	44.40	0.20	5.58	−0.24	−7.21	−0.08	−2.20	−0.31	−7.74	0.03	0.85	0.73	1512
Min	−0.04	−1.75	0.21	1.79	−0.07	−1.88	−0.65	−21.75	−0.34	−10.66	−0.79	−19.88	−0.09	−4.84	0.13	272
Max	0.14	2.41	1.29	104.38	0.53	17.40	0.11	2.47	0.20	1.41	0.05	0.34	0.25	12.46	0.92	5158

This table presents the results of a six-factor performance regression model via which the daily excess return of AI ETFs (calculated in Net Asset Values terms) is regressed on the excess return of the Russell 3000 Index, and the Fama and French (2015) SMB (Small Minus Big) factor, HML (High Minus Low book-to-price ratio) factor, CMA (Conservative Minus Aggressive) factor, RMW (Robust Minus Weak) factor, and Carhart’s MOM (momentum) factor. The estimation period of each AI ETF spans from its launch date up to 31 December 2025 (see Table 1 for launch dates). ^a indicates statistical significance at 1%; ^b indicates statistical significance at 5%; ^c indicates statistical significance at 10%.

reports the corresponding results when ETF returns included in the model have been calculated with Net Asset Values. By comparing the results in Table 6 and Table 7 we observe the proximity between the two sets of estimates. Therefore, the inferences drawn via analyzing trade price returns still hold in the case of NAV returns.

Overall, the finding that Fama and French and Carhart factors can explain a substantial portion of AI ETF performance has important implications for the interpretation of thematic investing. Our results indicate that the returns of both active and passive AI ETFs are largely driven by exposure to well-established systematic risk factors, including growth, value, profitability, investment, and momentum, rather than by the consistent creation of distinct thematic alpha associated with artificial intelligence innovation.

This finding suggests that AI ETFs predominantly act as vehicles that repackage conventional factor exposures within a thematic narrative, while still allowing for the possibility that some residual thematic effects may not be fully captured by standard asset pricing models. In this respect, AI-focused investment strategies exhibit considerable overlap with traditional factor-based investing, particularly with exposures to technology and growth-oriented firms.

From an investor perspective, this overlap implies that similar risk–return profiles could, to a significant extent, be replicated through diversified portfolios of broad-based ETFs at lower cost. At the same time, we acknowledge that thematic ETFs may offer convenience, targeted exposure, and narrative appeal that extend beyond pure factor replication.

More broadly, our findings underscore the importance of carefully distinguishing between genuine thematic alpha and returns attributable to established asset pricing factors, thereby contributing to a more nuanced evaluation of the economic value of thematic investment strategies.

7. Conclusions

Artificial Intelligence provides very prosperous opportunities for investors, and also assists financial managers in the securities selection process and portfolio management. The current study focuses on an investment product that has been attracting the interest of investors over recent years. This product is AI ETFs, that is, ETFs that invest in companies that are somehow involved in artificial intelligence.

In our analysis, we adopt an “active vs. passive” research stance. In doing so, we select 25 active and 22 passive AI ETFs with full trade data at least for 2025. Several issues are addressed, concerning the return and risk of AI ETFs, their pricing efficiency, or (in other words) the divergence between trade prices and NAVs of these ETFs, the persistence in these differences, and the impact of pricing inefficiency and intraday price volatility on returns. The relationship between AI ETFs’ performance and market factors regarding size, value, profitability, investment and momentum is evaluated too.

The empirical findings indicate that passive AI ETFs have outperformed their active peers as they have achieved considerably higher cumulative returns over their entire trade history. At the same time, the risk figures of the two groups approximate each other. Achieving significantly higher cumulative returns with the same level of risk constitutes a clear advantage of passive AI ETFs over active ETFs. On the other hand, active AI ETFs are better than passive ETFs at trading close to their underlying fundamental value. In any case, both ETF groups trade, on average, at a premium to their NAVs. The premium for both groups is found to be quite persistent for up to five trade days, possibly showing that investors could cash in on this premium persistence.

On the question of whether the premium is a material factor to consider when trying to determine the returns of AI ETFs, the empirical results reveal a positive relationship between the current premium and current returns, especially for passive AI ETFs. On the other hand, the lagged premium is negatively related to return, both for active and passive AI ETFs. Moreover, a negative relationship between return and concurrent intraday volatility is accentuated. This finding concerns both active and passive AI ETFs. A positive but weaker relationship between return and one-lagged intraday volatility is revealed too.

Finally, the regression analysis on the relationship between ETFs’ performance and the market factors finds that only a small number of AI ETFs can achieve positive and significant alphas over the Russell 3000 Index, which has been used as the market proxy index. In addition, positive relationships are found between the performance of AI ETFs and the size and momentum factors. The opposite is the case for the value and robustness factors. On the other hand, the effect of the conservativeness factor on ETFs’ performance is less significant.

Overall, our study provides useful empirical insights into a growing investment tool. To our view, the comprehensive study of AI ETFs’ performance, risk and pricing efficiency, as well as the accentuation of several factors that can be significant in determining the return of these ETFs should be quite helpful to a range of agents including ETF investors, assets managers and ETF issuers.