Financial Econometrics Beyond Estimation: Identification, Dependence, and Uncertainty

Abstract

Financial econometrics has expanded rapidly in recent decades, giving researchers many new tools for empirical analysis. Yet empirical findings often remain sensitive to modeling choices, sample construction, and specification decisions, suggesting that many disagreements arise not from estimation methods alone but from deeper limitations in what data can reveal. This survey organizes financial econometrics around three issues that determine the credibility of empirical inference: identification, dependence, and uncertainty. Identification concerns whether economic quantities such as causal effects, risk premia, and structural parameters can be credibly recovered from observable data. Dependence recognizes that assets, firms, and markets are interconnected, reducing the amount of independent information contained in financial data. Uncertainty captures not only sampling variation but also model misspecification, measurement error, and competing explanations. We review how these problems arise in asset pricing, corporate finance, ESG, and risk management and discuss implications for empirical design and inference.

1. Introduction

Financial econometrics has evolved into a methodologically diverse and computationally advanced field in recent decades. Primary drivers for this have been growth and development in estimation techniques, volatility modeling, and panel data methods (Campbell et al., 1997; Cochrane, 2005; Andersen et al., 2003). The availability of high-frequency data and recent advances in machine learning have further expanded the tools available to researchers in finance (Gu et al., 2020; Chen et al., 2024). Such methodological advances have facilitated increasingly precise investigations of asset pricing, corporate finance, risk management, and financial stability.

Despite these developments, empirical findings in finance often exhibit a lack of consistency and robustness across samples, methods, and modeling choices. Financial market anomalies tend to appear and disappear depending on time period, asset class, and testing methodology. Many causal claims in corporate finance rely crucially on identifying assumptions. Risk measures that perform well during stable periods frequently fail during times of market stress. These patterns suggest that differences in empirical outcomes result not from underlying economic theory alone but from deeper econometric challenges.

Review Scope and Methodology

Because the completeness and objectivity of any survey depend on how the underlying literature was identified and selected, we clarify our approach here. This review is conceptual and thematic in nature: its purpose is to organize existing methodological literature around a three-pillar framework rather than to catalogue every published study in financial econometrics through a formal systematic-review protocol. Candidate studies were identified through structured keyword searches of the Web of Science, Scopus, EconLit, and SSRN databases, using combinations of terms including “identification,” “endogeneity,” “dependence structure,” “tail risk,” “model uncertainty,” “robustness,” “machine learning,” “causal inference,” and “financial econometrics,” together with forward- and backward-citation tracking from foundational texts (Campbell et al., 1997; Cochrane, 2005) and from closely related handbook chapters and surveys (Roberts & Whited, 2013; Mukherjee et al., 2020; Kelly & Xiu, 2023).

The literature reviewed spans publications from the early 1980s, when the identification critique in econometrics was first articulated (Leamer, 1983), through 2024. Priority was given to two categories of work: methodologically foundational contributions that introduced an estimator, test, or identification strategy still in active use, and more recent contributions, concentrated after 2015, that extend these ideas to machine learning, high-dimensional, and network settings. Studies were included if they addressed identification, dependence, or uncertainty in an empirical finance context, either methodologically or through a widely cited application; purely theoretical asset-pricing papers without an empirical or econometric focus, and papers concerned only with estimation mechanics rather than identification, dependence, or uncertainty, were generally excluded. As with any thematic synthesis, the resulting review reflects our judgment about which strands of the literature are most consequential for applied credibility in finance, rather than an exhaustive enumeration of all published work; we return to this framework, and to how it differs from prior reviews, immediately below.

This survey proposes that many of these challenges rest on three interrelated deficiencies: identification, dependence modeling, and uncertainty quantification. Although identification has received the most sustained attention in the literature, all three elements jointly determine what can be learned from financial data, the assumptions required to learn it, and the confidence with which researchers can state their claims. Ignoring any one of them risks statistically precise yet economically ambiguous conclusions.

The proposition that estimation is not inference is not new. Leamer’s (1983) call to “take the con out of econometrics,” Manski’s (2003, 2007) work on partial identification, and Angrist and Pischke’s (2010) defense of credible research design in applied micro econometrics all argue, in different ways, that statistical precision is not evidence of validity. Within finance specifically, Roberts and Whited (2013) survey endogeneity problems in corporate finance, Lewellen et al. (2010) scrutinize the identification of risk-based explanations in asset pricing, and Kelly and Xiu (2023) synthesize machine-learning tools for empirical asset pricing while flagging their inferential limitations. Mukherjee et al. (2020) survey volatility and jump estimators, concentrating on the mechanics of estimation itself. These are valuable, focused contributions, but they largely treat identification, dependence, and model uncertainty as separate studies organized around specific techniques (an instrument, a volatility estimator, a network model). Instead, they should be treated as jointly operating constraints on inference.

Our contribution differs from earlier surveys in two important ways. First, it treats identification, dependence, and uncertainty not as independent technical concerns to be handled by separate toolkits, but as three interacting dimensions of a single inferential problem: weak identification is more consequential when dependence is strong, and dependence inflates apparent precision only because uncertainty is understated. Second, instead of simply listing and defining estimators, the survey organizes existing methodological literature. It spans traditional panel and instrumental variable methods as well as more recent causal machine learning, high-dimensional, and quasi-experimental techniques. We document what each of these factors contribute to, and what is left unresolved about these three pillars. The objective is a framework that generalizes across models, datasets, and applications, rather than a simple enumeration of techniques.

These three dimensions are not an arbitrary taxonomy; each maps onto a long-standing distinction in the methodological literature. Identification, in the sense developed by Manski (2003, 2007), concerns the conditions under which a parameter can be recovered at all from a data-generating process and a set of maintained assumptions. Dependence, in the sense developed by Engle (2002) and White (1980), concerns the conditions under which the effective information content of a sample departs from its nominal size. Uncertainty, in the sense developed by Hansen and Sargent (2008), concerns the conditions under which an estimator may converge to the wrong object, or to an unstable one, even when it is computed correctly. Table 1 situates the present survey against representative prior reviews in financial econometrics to make explicit how its scope and contribution differ; where earlier surveys are organized around a technique, an asset class, or a single dimension of the inference problem, this survey is organized around the interaction of all three.

Table 1. Comparison of This Survey with Selected Prior Reviews in Financial Econometrics.

Survey	Primary Focus	Identification	Dependence	Uncertainty	Gap Addressed by This Survey
Roberts and Whited (2013)	Endogeneity in corporate finance	Central; instruments, fixed effects, quasi-experiments	Not addressed directly	Not addressed directly	Extends identification concerns to dependence and uncertainty, and beyond corporate finance to asset pricing and risk management
Lewellen et al. (2010)	Identification of risk-based asset-pricing models	Central; cross-sectional test power and factor identification	Implicit, via correlated test assets	Limited; focused on statistical power	Generalizes a single asset-pricing critique into a framework applicable across finance subfields
Mukherjee et al. (2020)	Volatility and jump estimators under big data	Not a primary focus	Central, via high-frequency co-jumps	Estimator-level, not model-level	Connects estimator-level dependence measures to identification and model uncertainty
Kelly and Xiu (2023)	Machine learning in empirical asset pricing	Discussed as a limitation of ML tools	Addressed via factor structures	Discussed informally	Formalizes model uncertainty and dependence as pillars with explicit interactions rather than caveats
This survey	Cross-cutting framework for financial econometrics	Pillar 1, formalized and linked to the other two	Pillar 2, formalized and linked to the other two	Pillar 3, formalized and linked to the other two	Treats all three as jointly operating constraints, with an explicit interaction mechanism (Section 5) and applied checklist (Section 8)

The comparison is illustrative rather than exhaustive; it is restricted to reviews and highly cited methodological syntheses cited elsewhere in this survey. “Gap Addressed” summarizes, in our assessment, the principal respect in which the present framework extends each prior contribution.

Returning to the estimation—inference distinction introduced above, consider identification first. A researcher could estimate a parameter precisely and yet reach the wrong conclusion. Identification asks whether the data allow us to isolate the causal effect of the parameter of interest; if identification is weak, the regression coefficient may reflect the effect of omitted variables, measurement error, reverse causality, or equilibrium feedback. These issues are structural: they are built into how financial markets work. Asset prices reflect expectations about the future, firms make decisions based on those prices, and prices then adjust again in response to those decisions. The credibility of empirical finance arises less from using advanced techniques and more from the soundness of its assumptions. Sophisticated methods and large samples cannot substitute for a valid inferential framework. If credible identification is absent, then precision does not guarantee that the estimate represents any real economic effect.

Dependence problems in econometrics occur when observations are correlated rather than independent, across time, assets, or markets. While linear correlation captures some aspects of this co-movement, it falls short of many economically relevant features of financial data in times of market stress. Assets may move only mildly together in tranquil periods but much more strongly during crises; co-movement is often asymmetric and evolves dynamically (Engle, 2002; Patton, 2006). Dependence problems become particularly severe around extreme events, shaping tail risk and systemic stability (Longin & Solnik, 2001; Forbes & Rigobon, 2002). For a researcher, ignoring these forms of dependence can lead to misleading standard errors, distorted risk assessments, and spurious confidence in empirical results.

Finally, econometric challenges originating from uncertainty extend beyond finite-sample variability to model misspecification and the coexistence of competing explanations. When results change with the sample, it is often a symptom of limited or incomplete data rather than genuine instability in the underlying relationship. Small or moderate samples produce noisy estimates with wide confidence intervals; asymptotic theory may not approximate finite-sample reality well. Model risk increases under these conditions, producing conclusions that are fragile or consistent with multiple competing explanations. When uncertainty is inadequately quantified, empirical conclusions may appear sharper than warranted, masking sensitivity to assumptions, sample selection, and alternative model specifications (Hansen & Sargent, 2008).

Let us consider all three issues together in the well-studied relationship between financial market development and economic growth. To a researcher regressing financial market development on economic growth, it is unclear whether finance drives growth or growth drives finance; this simultaneity creates an identification problem, since ordinary least squares (OLS) estimates may be biased and fail to capture the true causal effect. This problem worsens if unobserved common factors, such as global economic shocks, affect multiple countries simultaneously; such correlated errors that represent a dependence problem, bias standard errors and undermine the reliability of statistical inference. And because growth-reducing channels such as credit misallocation are themselves noisy and hard to isolate, the measured relationship can look weaker than it is—an uncertainty problem layered on top of the other two. The three problems can reinforce one another. In the finance-growth example, simultaneity complicates identification, common shocks reduce the amount of independent variation, and noisy underlying mechanisms make the estimated relationship difficult to interpret.

This survey focuses less on collecting econometric techniques and more on understanding the limits of what financial data can tell us. Section 2 through Section 4 each examine one pillar, organized around four consistent questions: what theoretical advances define the pillar, what methodological efforts address it, what weaknesses and open problems remain, and where future research is most needed. Section 5 makes explicit how the three pillars interact. Section 6 illustrates these interactions through case studies in asset pricing, ESG research, and risk management. Section 7 discusses open challenges. Section 8 synthesizes the discussion into a summary table and a practical checklist for applied researchers. Section 9 concludes.

2. Identification in Financial Econometrics

A major constraint on empirical inference in finance is identification. It concerns whether variables of interest, such as risk premia, causal effects, and structural parameters, are uniquely recoverable from observable data under defensible assumptions. Estimation addresses how parameters are computed given a particular model; identification addresses whether the model and the data combined are informative enough to support a unique, economically interpretable answer. When identification fails, the issue is not merely that the economic interpretation is debatable, it is that the parameter itself is not statistically unique: the same data and model are consistent with more than one value of the parameter of interest.

2.1. Theoretical Foundations

Consider the simplest empirical framework, the regression:

$$Y_i=X_i\beta +u_i$$

(1)

with estimator:

$${\widehat{\beta }}_{OLS}={\left(X^{\prime }X\right)}^{-1}{X}^{\prime }Y$$

(2)

which is mechanically computable regardless of whether the data support a causal reading. Identification requires the orthogonality condition:

$$E\left[X_i{u}_i\right]=0$$

(3)

Estimation asks whether the coefficient can be computed from the data, whereas identification asks whether that coefficient represents the causal effect the researcher intends to measure. Even when Equation (3) holds and $\beta$ is point-identified, that requires the further assumption that $X_i$ is not itself a response to $Y_i$ or to a common cause of both. Point identification is therefore necessary but not sufficient for a causal interpretation.

Finance presents identification problems because financial markets data are often forward-looking. Prices, trading volume, and expectations are jointly determined, and observed correlations often reflect endogenous responses. Investors demand higher expected returns for perceived risk, but risk measurement is itself a function of prices and volatility, which are determined in equilibrium; risk and return are jointly set in markets. Many empirical disagreements in finance are not due to deficiencies in estimation technique but to ambiguity about what is actually identified. Improvements in estimation cannot compensate for identification deficiencies.

Several features of financial markets complicate identification relative to many other applied settings. First, simultaneity and equilibrium feedback are inherent to financial systems. Asset prices reflect expectations about the future, firms make decisions based on those prices, and prices then adjust again in response to those decisions, so what we observe may reflect an overall market equilibrium rather than a clear cause-and-effect relationship (Roberts & Whited, 2013). The assumption most likely to be violated is the orthogonality condition in Equation (3), the exogeneity assumption, which fails frequently because financial variables are endogenous by construction.

Second, controlled experiments in financial markets are scarce. Researchers often rely on natural experiments or regulatory changes, but these events usually affect several variables at once, which makes it harder to isolate a single causal effect. Even well-designed quasi-experiments can change expectations, market structure, or overall equilibrium behavior. For instance, the implementation of Regulation FD in 2000 is often used as a policy shock that altered the information environment; researchers compare outcomes before and after the rule to infer causal effects (Heflin et al., 2003; Mohanram & Sunder, 2006). Even though the “shock” is the rule banning selective disclosure, it indirectly affects analysts, investors, prices, liquidity, and corporate communication all at once, which is why identifying the causal effect of Reg FD on any one outcome is challenging.

Third, measurement errors and latent constructs are prevalent. Many central variables in finance such as risk, expectations, financial constraints, ESG performance are proxied rather than directly observed. Suppose a regressor is measured with error:

$$X_i^{obs}=X_i^{*}+{\eta }_i$$

(4)

Even if $E\left[X_i^{*}{u}_i\right]=0$, classical measurement error implies

$$E\left[X_i^{obs}{u}_i\right]\ne 0$$

(5)

weakening identification and biasing inference. When measurement error interacts with equilibrium responses, identification can become substantially weaker than standard diagnostics suggest.

Fourth, dependence affects identification. If observations are strongly correlated over time, across firms, or across markets, what looks like independent variation may reflect common shocks rather than exogenous variation in $X$. Even if the data appear to contain sufficient variation to identify the effect of $X$ on $Y$, much of that variation may be common-shock-driven rather than independent information, leading researchers to overestimate how well the effect is identified. We return to this interaction in Section 5.

2.2. Methodological Approaches

Empirical finance employs a range of identification strategies, each involving trade-offs between credibility and interpretability.

Panel data methods exploit within-entity variation to control for unobserved heterogeneity. Consider the basic fixed effects model:

$$Y_{it}=\beta X_{it}+{\alpha }_i+u_{it}$$

(6)

Fixed effects estimation differences out ${\alpha }_i$ using deviations from each unit’s mean. However, if a time-varying factor $Z_{it}$ affects both $Y$ and $X$, so that:

$$Y_{it}=\beta X_{it}+\gamma Z_{it}+{\alpha }_i+u_{it}$$

(7)

fixed effects cannot remove $Z_{it}$ because it changes over time; if $Z_{it}$ is correlated with $X_{it}$, then $\beta$ is biased. Similar issues arise when lagged outcomes $Y_{it-1}$ affect current $X$, creating dynamic feedback that fixed effects does not automatically resolve.

Fixed effects remove any influence that is constant for a given firm or asset over time, such as its industry or headquarters location, by comparing each unit only to its own average, which is what Equation (6) formalizes. Equation (7) shows that this fix breaks down whenever an additional factor that changes over time, not just across units, drives both the outcome and the regressor of interest.

Instrumental variables illustrate the distinction between estimation and identification directly. Consider the system:

$$Y_i=X_i\beta +u_i$$

(8)

$$X_i=Z_i\pi +v_i$$

(9)

Identification requires both instrument relevance:

$$\pi \ne 0$$

(10)

and the exclusion restriction:

$$E\left[Z_i{u}_i\right]=0$$

(11)

When $\pi$ is small, conventional asymptotic approximations become unreliable even with large samples; narrow confidence intervals in this setting do not imply that identification is strong.

Quasi-experimental designs like difference-in-differences (DiD), regression discontinuity designs (RDD), and synthetic control methods, have become the workhorses of applied identification in corporate finance and increasingly in asset pricing and banking research. RDD exploits a known threshold rule (e.g., an index inclusion cutoff, a regulatory asset-size threshold, a credit-score cutoff for loan approval) and compares outcomes for observations just above and below it, under the assumption that agents cannot precisely manipulate their position relative to the threshold (Lee & Lemieux, 2010). DiD compares changes in outcomes for treated and untreated units around an event, relying on a parallel-trends assumption; recent advances address the bias that arises when treatment timing is staggered across units, which earlier two-way fixed-effects estimators handled poorly (Callaway & Sant’Anna, 2021). Synthetic control methods construct a weighted combination of untreated units that closely tracks the treated unit’s pre-treatment outcome path, then use the gap after treatment as the causal estimate; this is particularly useful in finance when the “treated” unit is a single market, country, or large institution for which no natural individual control group exists (Abadie, 2021).

High-dimensional and causal machine-learning methods address a different identification problem: when there are many potential confounders and the researcher does not know, a priori, which of them belong in the model. Regularized selection methods choose controls in a way that is valid for inference on a low-dimensional causal parameter even when the set of candidate controls is large relative to the sample (Belloni et al., 2014). Double/debiased machine learning extends this idea by using flexible machine-learning estimators for the nuisance components of the model (the relationship between covariates and the outcome, and between covariates and the treatment) while preserving valid inference on the causal parameter of interest, through a construction that is insensitive to small errors in those nuisance estimates (Chernozhukov et al., 2018). These methods do not solve the exclusion-restriction or exogeneity problem; they solve a complementary problem of controlling for high-dimensional confounding without overfitting.

These high-dimensional and machine-learning tools come with their own limitations, which are easy to understate in a brief description. Regularized and tree-based methods typically sacrifice interpretability for flexibility: the fitted relationship between covariates and outcomes rarely has a closed form that a reader can inspect, so the economic mechanism behind a result must be argued separately from the statistical fit itself. Many of these methods are also computationally demanding to implement correctly in inference settings, since valid confidence intervals for the causal parameter typically require repeated sample-splitting or cross-fitting rather than a single model fit, and their finite-sample performance in the short, strongly dependent panels typical of finance is less well understood than their performance in the larger, more independent datasets common in labor and public economics. Bayesian and quasi-Bayesian variants of these methods can, in principle, propagate model-selection uncertainty into the reported interval, but this comes at the cost of prior sensitivity, discussed further in Section 4.2. In practice, the choice between a transparent linear specification and a flexible machine-learning specification in finance is therefore not simply a choice about statistical accuracy; it is a trade-off between interpretability, computational tractability, and robustness to the dependence structure of the data.

In financial econometrics, identification is rarely complete; it is more often partial, local, or identified only under strong assumptions. Sometimes the data tell us only that the parameter lies within a set or range rather than a single value (Manski, 2003; Imbens & Manski, 2004):

$$\beta \in \mathcal{B}$$

(12)

where $\mathcal{B}$ denotes the set of parameter values consistent with observed moments. Small changes in maintained assumptions can shift this set substantially, yet standard reporting usually shows a single point estimate and a p-value, obscuring how fragile the underlying identification really is.

2.3. Weaknesses and Open Problems

Asset pricing provides an established illustration of these weaknesses. Different models can often explain stock returns about equally well, so statistical fit alone does not resolve whether returns are high because of compensation for risk or because of mispricing (Lewellen et al., 2010). In corporate finance and ESG research, identification challenges are compounded by measurement error, endogeneity, and cross-sectional dependence: quasi-experimental designs yield valuable insights, but results frequently hinge sensitively on instrument choice, sample construction, and specification decisions. RDD and DiD designs, while more credible than naive OLS, are not immune to these problems. The parallel-trends assumption in DiD is itself untestable in the post-treatment period, and RDD estimates are local to the threshold, so they may not generalize to agents far from the cutoff. High-dimensional and causal machine-learning methods, in turn, inherit whatever confounding is not captured in the observed covariates; they improve efficiency and reduce overfitting bias but cannot manufacture exogenous variation where none exists.

These weaknesses are not merely additive; the literature does not always agree on how much they matter in a given setting. Proponents of quasi-experimental designs emphasize that RDD and staggered DiD are credible precisely because their identifying assumptions are explicit and partially testable (Lee & Lemieux, 2010; Callaway & Sant’Anna, 2021), while critics note that the local, threshold-specific nature of these estimates means a well-identified effect can still fail to generalize to the broader population a policymaker cares about; both claims are correct, but they answer different questions, and confusing them is itself a source of overstated credibility. Similarly, causal machine-learning methods are sometimes described as resolving the identification problem in high-dimensional settings, when a more accurate characterization is that they resolve a narrower variable-selection problem while leaving the exogeneity of the treatment itself exactly as assumption-dependent as before.

To summarize, several identification failures recur across applications: equilibrium endogeneity, instrument proliferation, weak identifying variation even in large datasets, measurement-driven results, and sensitivity to auxiliary assumptions.

2.4. Future Research Directions

Three directions stand out. First, greater integration of quasi-experimental designs (RDD, staggered DiD, synthetic control) with explicit sensitivity analysis would make the fragility of identifying assumptions more visible to readers, rather than leaving it implicit in robustness-table footnotes. Second, causal machine-learning methods are still underused in finance relative to labor and public economics; extending double machine learning and related approaches to settings with strong cross-sectional and temporal dependence, which is the norm in financial panels, remains largely open, and is revisited in Section 5. Third, set-identification approaches (reporting a defensible range for $\beta$ rather than a single point estimate) deserve wider adoption in applied finance, particularly where instrument strength or parallel-trends assumptions are contestable. To do credible empirical finance, researchers must clearly state their identifying assumptions, test how sensitive results are to those assumptions, and separate statistical precision from economic meaning. Identification is more than a step before estimation; it determines whether the findings have meaningful economic interpretation.

3. Dependence Beyond Correlation

Dependence is pervasive in financial data. Asset returns co-move across markets, firms respond to common macroeconomic and financial shocks, and risks materialize jointly, particularly in periods of stress. Yet in much empirical work, dependence is treated as a nuisance to be corrected with heteroskedasticity- or cluster-robust standard errors rather than examined as an economic object in its own right. This treatment understates its central role: dependence shapes identification, governs the effective information content of a sample, and fundamentally constrains inference.

3.1. Theoretical Advances

In empirical finance, we often start by assuming that observations are mostly independent, or at most loosely related. To measure how two variables move together, we typically use the correlation coefficient:

$${\rho }_{XY}={\displaystyle \frac{\mathrm{Cov}\left(X,Y\right)}{\sqrt{\mathrm{Var}\left(X\right)\mathrm{Var}\left(Y\right)}}}$$

(13)

While correlation captures average co-movement, it does not capture how a relationship changes across states of the world, whether it is asymmetric, or whether extreme events co-occur. For example, higher debt might be only moderately correlated with stock volatility on average, but during a financial crisis the relationship strengthens considerably as highly leveraged firms crash faster; correlation, being an average, misses this tail dependence.

Several stylized features of financial dependence are well documented. A first departure from simple correlation is nonlinearity: assets often show only modest correlation in tranquil periods but sharp co-movement during downturns. What matters economically is not average covariance but the likelihood of joint extreme losses. Tail dependence formalizes this: the upper tail dependence coefficient is:

$${\lambda }_U={\mathrm{l}\mathrm{i}\mathrm{m}}_{u\to 1^{-}}P\left(Y>F_Y^{-1}\left(u\right)\mid X>F_X^{-1}\left(u\right)\right)$$

(14)

where $F_Y^{-1}$ denotes the marginal quantile function. Two assets can exhibit low ${\rho }_{XY}$ yet high ${\lambda }_U$, implying limited diversification precisely when it is needed most (Longin & Solnik, 2001).

The question this coefficient is built to answer is a narrow one: among the worst outcomes for asset X, how often is asset Y also at its worst? A correlation computed across the full sample can miss this entirely, since it averages calm and crisis periods together.

This nonlinearity is also typically asymmetric—downside co-movement tends to exceed upside co-movement, driven by leverage effects, margin constraints, funding pressures, and investor behavior under stress. A single symmetric correlation coefficient averages the tranquil and crisis states and therefore understates downside risk, distorting portfolio diversification and capital structure decisions where the relevant danger lies in crisis behavior rather than normal-times behavior (Patton, 2006).

Because dependence strengthens in crises and weakens in tranquil periods, it is also time-varying rather than fixed. Dynamic conditional correlation models formalize this time-varying process through recursions such as:

$$Q_t=\left(1-\alpha -\beta \right)\bar{Q}+\alpha {\epsilon }_{t-1}{{\epsilon }_{t-1}}^{\prime }+\beta Q_{t-1}$$

(15)

where $Q_t$ governs time-varying covariance dynamics (Engle, 2002). Even dynamic models can smooth crises more than reality warrants, since they assume a specific functional form that may mischaracterize behavior in extreme states.

A further complication is scale: financial systems consist of many interconnected entities, so dependence is a high-dimensional problem as well as a nonlinear, asymmetric, and time-varying one. A parsimonious representation is the factor structure:

$$R_{it}={{\beta }_i}^{\prime }{F}_t+{\epsilon }_{it}$$

(16)

where common factors $F_t$ drive co-movement. Determining how many such factors are needed, rather than assuming the number a priori, is itself a nontrivial inferential problem; consistent, data-driven criteria exist for recovering the number of factors in large panels (Bai & Ng, 2002). Factor models reduce dimensionality dramatically but do not capture direct firm-to-firm spillovers, as when one bank’s failure directly harms specific counterparties regardless of broad market factors.

Each asset’s return in Equation (16) decomposes into a shared, market-wide component, driven by common factors that affect many assets at once, and an asset-specific component that is, ideally, unrelated to any other asset’s idiosyncratic return. The practical difficulty is that the number of common factors, and therefore how much of the co-movement in the data is ‘shared’ versus ‘independent,’ is not known in advance and must itself be estimated.

3.2. Modeling Dependence

Copula-based approaches separate the marginal distributions of assets from their dependence structure, accommodating tail behavior and asymmetry that correlation cannot (Patton, 2006). Multivariate volatility models such as DCC allow time-varying covariance structures (Engle, 2002). Network models treat firms as nodes connected by economically meaningful links rather than assuming co-movement is driven only by broad market factors: if Bank A lends heavily to Bank B and Bank B fails, Bank A suffers losses directly, even where an aggregate market factor would not fully explain it (Billio et al., 2012; Diebold & Yılmaz, 2014). High-frequency data have also enabled realized-measure approaches to dependence. Computing daily realized correlations and jumps from intraday prices rather than relying on lower-frequency proxies can reveal predictive relationships between uncertainty and co-movement that daily-frequency correlation measures obscure (Gkillas et al., 2021). Each of these frameworks is more sophisticated than a single correlation coefficient, but each requires modeling choices. Functional forms, regularization parameters, and threshold rules influence conclusions and introduce a further layer of uncertainty. The field has therefore moved from “too simple” to “more sophisticated but model-dependent,” rather than from “uncertain” to “certain.”

Dependence has immediate implications for inference. Standard errors and hypothesis tests are commonly derived under assumptions of independence or weak dependence. When serial dependence is present, effective information content falls. A rough approximation is:

$$n_{\mathrm{eff}}\approx {\displaystyle \frac{n}{1+2{\displaystyle {\sum }_{k=1}^{\mathrm{\infty }}}{\rho }_k}}$$

(17)

where ${\rho }_k$ denotes autocorrelation at lag $k$. Strong persistence implies that the nominal sample size overstates effective variation. Cross-sectional dependence produces a similar distortion, as common shocks generate apparent statistical significance across firms or assets that does not reflect independent evidence.

3.3. Weaknesses and Open Problems

Tail dependence is particularly consequential for financial risk and systemic stability, since variance-based risk measures are poorly equipped to capture joint extremes; during crises, tail dependence intensifies, undermining diversification benefits precisely when they are most relied upon (Longin & Solnik, 2001). Measures such as CoVaR quantify joint tail exposure but depend critically on specification and estimation assumptions (Adrian & Brunnermeier, 2016). More fundamentally, dependence interacts with identification: in panel settings with strong cross-sectional correlation, within-entity variation may reflect shared macroeconomic movements rather than independent shocks, so identification arguments that implicitly assume independence across units can be substantially weaker than they appear. Clustering standard errors adjusts inference for a known dependence structure but does not resolve this deeper problem, since it corrects the standard error of an estimate whose identifying variation may itself be compromised by the same dependence.

The literature is similarly divided on how much of this dependence is already handled by standard practice. Clustering standard errors is often treated as a sufficient fix once a plausible clustering dimension is identified, yet clustering only corrects the standard error of an estimate; it does not verify that the identifying variation used to obtain that estimate is itself independent of the dependence structure being clustered on. Studies that report similar point estimates under alternative clustering schemes are frequently taken as evidence of robustness, but this practice can mask the more fundamental concern raised throughout this section, that the underlying variation may be common-shock-driven regardless of how the standard errors are computed.

3.4. Future Research Directions

Progress requires recognizing dependence as an economic feature of financial markets. It should not be treated like a technical nuisance that can be swept into a robust standard error. Promising directions include network-based identification strategies that use the structure of financial linkages themselves as a source of exogenous variation, high-frequency-based dependence measures that can detect regime shifts in co-movement before they are visible at lower frequencies (as in Gkillas et al., 2021), and closer integration between dependence modeling and the high-dimensional causal-inference methods discussed in Section 2, since both must ultimately confront the same underlying interconnectedness of financial data. A research design that is aware of dependence issues, costs little at the outset and prevents the more expensive failure of discovering, after publication, that an identification strategy’s power came from correlated shocks and not independent variation.

4. Uncertainty, Robustness, and Model Risk

Financial data are often characterized by uncertainty. These data are often noisy, strongly dependent, and generated by strategic, forward-looking agents operating within changing institutional environments. Yet researchers often focus on precise point estimates $\widehat{\theta }$ and test statistics, which can convey a false sense of accuracy. A sound approach to financial econometrics must treat uncertainty as a fundamental limitation on what can realistically be learned from the data, not merely as variation around a point estimate.

4.1. Theoretical Formulation

In standard econometric formulations, inference proceeds from an estimator:

$${\widehat{\theta }}_n={\theta }_0+{\displaystyle \frac{1}{\sqrt{n}}}{Z}_n+o_p\left(n^{-1/2}\right)$$

(18)

where ${\theta }_0$ is the true parameter and $Z_n$ converges in distribution to a normal random variable. The estimate is a function of the true value plus random error that shrinks as the sample grows. This representation highlights sampling uncertainty, which persists even under correct specification because finite samples generate dispersion around ${\theta }_0$. In financial data, observations are not truly independent, so dependence limits the effective sample size $n_{\mathrm{eff}}$ to be far smaller than the nominal $n$ (as in Equation (17)). If $n_{\mathrm{eff}}<n$, the variance of ${\widehat{\theta }}_n$ scales with $1/n_{\mathrm{eff}}$ rather than $1/n$, so the estimate appears more precise than it really is; apparent precision may therefore reflect unrecognized dependence rather than genuine independent information.

Put simply, Equation (17) formalizes an intuitive idea: if today’s observation is highly predictable from yesterday’s (strong serial dependence, large ρ, k), then each new observation adds less genuinely new information than an independent observation would, so the effective sample size n can be much smaller than the number of rows in the dataset, N. A standard error computed as if n equaled N will therefore be too small, making the result look more precise than the data actually support.

Estimation uncertainty is compounded by the statistical properties of financial returns, which frequently exhibit heavy tails and excess kurtosis, meaning extreme gains and losses occur more often than standard models assume. In a linear pricing relation:

$$r_{it}={\alpha }_i+{{\beta }_i}^{\prime }{f}_t+{\epsilon }_{it}$$

(19)

a small number of extreme observations can disproportionately influence estimates of beta or risk premium. Asymptotic normal approximations may not work well at realistic sample sizes under heavy tails, and standard errors can understate the true variability of the estimate. This problem is exacerbated because extreme returns often cluster in time. For example, during financial crises effective information is reduced, and estimation variance is inflated beyond what independent-sampling theory would suggest.

4.2. Quantifying Uncertainty

In empirical finance, uncertainty does not come only from sampling or estimation; it also comes from model uncertainty. Researchers often face several models that fit the data almost equally well, yet each implies different economic interpretations. Choosing among competing factor structures, identification assumptions, or functional forms—including, for instance, the choice of how many latent factors to retain in a factor model (Bai & Ng, 2002)—is not resolved simply by picking the specification with the best in-sample fit, since that choice can itself be arbitrary; model selection therefore becomes a source of uncertainty in its own right.

The issue is compounded when the model itself is mis-specified—when the true data-generating process lies outside the class of models under consideration. Suppose the true process lies outside the maintained model class, so that the estimator converges to:

$${\theta }^{*}=\mathrm{a}\mathrm{r}\mathrm{g}{\mathrm{m}\mathrm{i}\mathrm{n}}_{\theta }E\left[m{\left(X_t,\theta \right)}^2\right]$$

(20)

rather than satisfying the population moment condition $E\left[m\left(X_t,{\theta }_0\right)\right]=0$. Estimators then converge to a pseudo-true parameter—the value that best approximates the data within the chosen model, even though it may not represent the true underlying economic relationship. Precision around ${\theta }^{*}$ reflects stability within a potentially misspecified framework, not correctness.

What this describes, in practical terms, is a model that is simply wrong in some respect such as an omitted nonlinearity or a missing state variable. The estimator does not fail outright; it still converges to a well-defined number as the sample grows. But that number, the “pseudo-true” parameter, is only the best approximation available within an incorrect model class, and need not equal the true economic quantity the researcher set out to estimate. A tight confidence interval around a pseudo-true parameter is still tight, but it is tight around the wrong target.

Because of these uncertainties, robustness and sensitivity analysis are important parts of good empirical research. More broadly, robust inference has long sought to reduce sensitivity to violations of classical assumptions, beginning with heteroskedasticity-consistent covariance estimation (White, 1980) and extending to broader concerns regarding model uncertainty and misspecification. Robustness analysis checks whether a result holds when reasonable aspects of the model change, for example different specifications, samples, or control variables. If the estimated effect changes substantially across these variations, that instability itself signals uncertainty about the result (Hansen & Sargent, 2008). Sensitivity analysis goes further, deliberately varying key assumptions like instrument validity, exclusion restrictions, or how dependence is modeled and then examining how much the results change. Instead of reporting a single estimate $\widehat{\theta }$, researchers can report a range of values consistent with the data and assumptions, shifting the focus from a precise point estimate to bounds on plausible values (Manski, 2007); in financial contexts with weak instruments, short time series, or strong dependence, bounded conclusions can be more credible than precise but fragile ones.

Different methodological paradigms confront these challenges in distinct ways. Bayesian approaches represent uncertainty through posterior distributions $p\left(\theta \mid X\right)$, integrating prior beliefs with observed data and, in principle, allowing a formal treatment of model uncertainty through model averaging; yet when identification is weak or data are limited, posterior inference can be sensitive to prior choices. Frequentist approaches focus on how estimates would behave under repeated sampling and have well-understood statistical properties, but because they rely on asymptotic approximations, they may understate uncertainty in complex, dependent environments. The key distinction is not really Frequentist versus Bayesian methods; it is between approaches that make uncertainty explicit and those that conceal it behind significance tests or conventional thresholds, such as a 5% p-value.

This trade-off is not merely theoretical. Bayesian model averaging and shrinkage priors can, in principle, formalize the model uncertainty discussed above by placing a probability distribution over competing specifications rather than selecting a single one; but in the short, strongly dependent samples typical of finance, posterior model weights can be highly sensitive to the choice of prior, particularly when several models fit the data nearly equally well, which is precisely the situation in which model averaging is most needed. Fully Bayesian treatments of high-dimensional financial panels are also computationally demanding, since posterior sampling over many candidate models or many candidate controls scales poorly, which has limited their routine use outside of specialized applications. These are not reasons to avoid Bayesian methods, but they mean that a posterior interval should not automatically be read as a more complete accounting of uncertainty than a frequentist one; both approaches face a common constraint, that the data available in finance are rarely rich enough, on their own, to overcome strong prior or asymptotic assumptions.

4.3. Weaknesses and Open Problems

A source of uncertainty specific to the research process itself, rather than to any single dataset, is often overlooked: researcher degrees of freedom, specification searching, and publication bias. When many plausible specifications are tried and only the strongest results are reported or published, conventional significance thresholds substantially overstate the credibility of a finding. This concern is not hypothetical in finance: reviews of the cross-sectional return-predictability literature find that most published “anomalies,” when subjected to a multiple-testing correction appropriate to the hundreds of factors that have been tested over time, no longer clear a defensible significance bar (Harvey et al., 2016), that published-factor returns tend to decay substantially after publication, consistent with either statistical overfitting or price adjustment to newly public information (McLean & Pontiff, 2016), and that a large share of documented anomalies fail to replicate once data-screening choices such as the treatment of microcap stocks are standardized (Hou et al., 2020). These findings illustrate that uncertainty in financial econometrics extends beyond any individual study’s sampling error to the collective research process that generates and screens findings before they reach publication.

Uncertainty of this kind is closely related to broader concerns about computational reproducibility that have received growing attention across empirical social science. Even when a specification is not deliberately searched over, results in financial econometrics often prove difficult to reproduce exactly because of undocumented data-cleaning choices, proprietary or vendor-specific data vintages, and code that is not shared alongside the published article; two researchers applying “the same” method to “the same” underlying data can obtain materially different estimates for reasons that have nothing to do with sampling variation. The broader replication crisis documented in psychology and other empirical fields has direct analogues in finance, where the anomaly-replication exercises discussed above (Harvey et al., 2016; McLean & Pontiff, 2016; Hou et al., 2020) function as a finance-specific version of the same concern. Journals and researchers increasingly request that code and data be archived at publication, and some finance journals now require replication packages, but adoption remains uneven, and even a complete replication package does not, by itself, reveal how many alternative specifications were considered before the archived one was selected.

4.4. Future Research Directions

Financial econometrics that studies identification and systemic dependence must recognize multiple, simultaneous sources of uncertainty: sampling variability, extreme observations, dependence, competing models, model misspecification, and the research process itself. Promising directions include wider use of pre-registration and multiple-testing corrections calibrated to the number of specifications a researcher has effectively searched over (extending Harvey et al., 2016), systematic replication exercises in the spirit of Hou et al. (2020) applied to newer machine-learning-based findings, and further development of Bayesian model-averaging tools that remain tractable in high-dimensional, strongly dependent financial panels. A concrete step available to authors, journals, and reviewers today is to require, alongside a replication package, a short specification-curve or multiverse-style appendix showing how the headline estimate moves across a defensible range of alternative data-cleaning, sample-construction, and control choices, rather than leaving readers to infer robustness from a handful of hand-picked robustness checks. The goal is not to eliminate uncertainty, which is impossible, but to measure and communicate it more clearly.

5. The Interdependence of Identification, Dependence, and Uncertainty

Section 2, Section 3 and Section 4 treated identification, dependence, and uncertainty separately for clarity, but they are not independent problems that happen in finance concurrently. Each shapes the severity of the other two. Figure 1 summarizes the three principal channels of interaction developed throughout this survey.

Figure 1. How the three pillars reinforce one another. The three pillars of empirical credibility in financial econometrics reinforce one another. Strong dependence erodes the independent variation needed for identification, weak identification widens the range of models consistent with the data, and unmodeled dependence understates true sampling uncertainty.

Dependence weakens identification directly. Identification strategies rely on a source of variation in the regressor that is plausibly unrelated to the outcome except through the channel being studied—but when observations are strongly dependent, as when firms share exposure to common shocks or time periods share persistent macroeconomic conditions, that variation is often itself a reflection of the same shocks affecting the outcome. The identifying assumption is consequently weaker than it appears (Section 2.1, point four; Section 3.3).

This has a direct consequence for uncertainty: when a parameter is only partially identified, the set of models consistent with the data, $\mathcal{B}$ in Equation (12), is itself wide, and this widening compounds ordinary sampling and model uncertainty (Section 4.2) rather than substituting for it—a precise point estimate reported in its place understates two sources of uncertainty at once. Dependence compounds the problem further on its own terms: because conventional standard errors are typically derived under independence or weak dependence, unmodeled dependence deflates the effective sample size, $n_{\mathrm{eff}}$ in Equation (17), which mechanically inflates true sampling uncertainty even when the reported standard error gives no hint of it (Section 4.1).

The practical implication is that these three pillars cannot be addressed one at a time and independently. First identify, then adjust standard errors for dependence, then report a confidence interval without the risk of understating the joint uncertainty surrounding a finding. A credible empirical strategy in finance should instead ask, for any single result: what identifying assumption is being invoked, how much of the identifying variation could plausibly reflect dependence rather than independent variation, and how would the conclusion change under a defensible range of alternative specifications. Section 6 illustrates this joint reasoning in three applied settings, and Section 8 translates it into a practical checklist.

6. Cross-Cutting Implications and Case Illustrations

The interaction of identification, dependence, and uncertainty becomes most evident in applied settings where econometric assumptions meet the complexities of real financial data. These issues are often discussed separately in methodological work, while in practice they appear jointly and shape the interpretation of empirical findings.

Asset pricing: Numerous return factors have been proposed over time, many of them highly correlated with one another, unstable across samples, and sensitive to how portfolios are constructed, which makes it difficult to identify which factors genuinely drive returns. The scale of this problem has become clearer as the literature has confronted it directly: applying a multiple-testing correction appropriate to the number of factors that have actually been tested against the data raises the bar for statistical significance considerably, and only a small fraction of previously published factors clear it (Harvey et al., 2016). Consistent with data mining rather than genuine discovery, many published factors also show a marked decline in average returns after the year they are first documented (McLean & Pontiff, 2016), and systematic replication exercises that standardize data-construction choices such as the exclusion of microcap stocks eliminate the significance of a large share of previously “confirmed” anomalies (Hou et al., 2020). This is precisely the interaction described in Section 5: the identification problem (which factor, if any, reflects a real risk exposure) is inseparable from the uncertainty problem (how many specifications were implicitly searched over before publication) and from the dependence problem (returns on competing factors are themselves highly correlated, so apparently independent “confirmations” share common variation). In such settings, empirical tests do not simply evaluate a single hypothesis; they implicitly choose among many plausible models and specifications, most of which were never reported.

A concrete illustration makes the mechanism explicit. Suppose a newly proposed factor earns a statistically significant average return in a sample ending in year t. If that estimate is one of several hundred similar factor constructions effectively tested across the literature (the uncertainty problem), and if its returns are highly correlated with several previously published factors because all are proxying for a similar underlying source of common variation (the dependence problem), then the marginal evidence that this specific factor reflects a distinct risk exposure, rather than a relabeling of existing exposures or a product of selective reporting, is far weaker than a single-study p-value would suggest (the identification problem). Distinguishing a genuinely new risk factor from a repackaged version of an existing one therefore requires evidence on all three pillars simultaneously, not a single significance test.

ESG research: ESG scores are furnished by different providers using proprietary methodologies to aggregate diverse firm characteristics, so firms can receive very different scores depending on the provider (Berg et al., 2022; Gibson Brandon et al., 2021). For researchers, this creates both measurement uncertainty (which score is the “correct” one) and identification concerns (is a documented ESG-performance relationship driven by the underlying construct or by the scoring methodology). At the same time, firms are jointly exposed to common factors such as regulatory changes, macroeconomic conditions, and industry-specific shocks, generating strong cross-sectional dependence. As a result, studies of the ESG-performance link often find that results shift depending on how ESG scores are constructed, which firms are included in the sample, and which controls are used. As in the asset pricing case, the identifying variation is inseparable from the broader dependence structure of the data.

To see why this matters for a specific empirical claim, consider a study reporting that higher ESG scores predict lower stock return volatility. If the ESG score used is one particular vendor’s construction, the same relationship can reverse sign or lose significance when a differently constructed score from another vendor is substituted for it, since the vendors disagree substantially in how they weight the same underlying firm disclosures (Berg et al., 2022). If, in addition, high-ESG firms cluster in particular sectors that were also less exposed to a common shock over the sample period, the estimated relationship may partly reflect that shared exposure rather than any effect of ESG performance itself. A finding that survives multiple vendor scores, multiple sample periods, and industry-adjusted comparisons is far more credible than one that relies on a single score and a single sample, yet the latter remains far more common in the published literature.

Risk management: Many widely used risk models are calibrated on historical data drawn from relatively stable periods. When financial conditions change abruptly, correlations between assets can increase sharply, and extreme losses can occur together, so parameters estimated from stable periods no longer reflect the true level of risk. The performance of pre-crisis value-at-risk models during the 2007–2009 financial crisis is a well-documented example: many such models substantially underestimated joint downside risk, contributing to excessive confidence in diversification and capital adequacy that proved costly once dependence structures shifted (Acharya, 2009; Adrian & Brunnermeier, 2016). The above example illustrates all three pillars simultaneously; model uncertainty about which volatility and correlation model applied in a crisis regime, dependence that intensified precisely when the models assumed it would not, and an identification problem in attributing losses to a specific risk factor once assets began moving together for reasons unrelated to their individually estimated exposures.

The practical cost of overlooking these interactions was not merely academic. Capital buffers, margin requirements, and portfolio hedges calibrated on pre-crisis correlation and volatility estimates were, in effect, calibrated on an implicit model of dependence that the crisis itself invalidated, so institutions that appeared well-diversified under the pre-crisis model discovered that their effective diversification had been far smaller than reported once assets began moving together. This example illustrates why treating identification, dependence, and uncertainty as separate technical checkboxes, rather than as jointly binding constraints, can translate directly into mismeasured risk with real financial consequences.

Across these three settings, a common lesson can be learned: a divergence in ESG scores, a spike in crisis correlations, and a decayed factor return are not separate methodological issues but the same underlying weaknesses—identification, dependence, and uncertainty compounding each other, showing up in whichever pillar happens to be most visible in that literature.

7. Open Challenges and Future Directions

Financial econometrics has made substantial progress, but the focus has largely been on statistical technique rather than on how identification, dependence, and uncertainty interact in complex financial environments. The main challenge going forward is not a shortage of estimators but the difficulty of combining these elements within models that realistically reflect how financial markets work.

One important issue is integrating identification strategies with realistic models of dependence. Much empirical work implicitly assumes that the variation used to identify causal effects is orthogonal to the dependence structure of the data. In financial contexts, the variation used for identification, whether it is an instrument, policy shock, or discontinuity, often arises from common shocks, industry cycles, or network interactions among firms and markets. When such dependence is ignored, both identification arguments and inferential procedures become fragile.

A second challenge is drawing reliable inference from many dependent variables. The growth of financial datasets has encouraged machine learning and regularization methods for prediction and variable selection. These approaches work well for prediction, but their inferential properties in strongly dependent environments remain less well understood; in financial applications, where the objective often extends beyond prediction to economic interpretation, developing methods that quantify uncertainty reliably in high-dimensional, dependent data remains an open problem (Belloni et al., 2014; Chernozhukov et al., 2018).

Two further strands of methodology lie outside the scope of the present framework but bear directly on the three pillars. Quantile and distributional treatment-effect methods characterize how identification, dependence, and uncertainty vary across the outcome distribution rather than only at its mean, which is particularly relevant for tail risk and downside co-movement (Section 3). Market microstructure models of price formation at the highest frequencies confront identification and dependence at time scales where the equilibrium-feedback problems discussed in Section 2 are most acute. Extending the three-pillar framework to these settings is a natural direction for future research.

A related challenge is identification under model uncertainty. In finance, multiple models can explain many phenomena while fitting the data similarly well. Frameworks that explicitly acknowledge such uncertainty—set identification, model averaging, and systematic sensitivity analysis—help reveal the range of conclusions the data support (Manski, 2007) but remain underused in applied financial research.

More concretely, three actionable directions stand out for researchers deciding where to focus new work. First, financial network data on interbank exposures, supply-chain linkages, and institutional ownership overlap, remain underused as a source of identifying variation in their own right, not only as an object of study; using the structure of a network (for example, which firms are connected only indirectly to a shock) to isolate exogenous variation is a concrete, underexploited alternative to relying solely on policy shocks. Second, artificial-intelligence-based text and alternative-data methods (parsing earnings calls, regulatory filings, or news flow) generate high-dimensional, strongly dependent predictors whose implications for identification and effective sample size, in the sense of Equation (17), have not been systematically worked out; extending the dependence-robust inference tools discussed in Section 3 to these new data sources is a well-defined, tractable research agenda rather than a vague call for “more data.” Third, market microstructure settings, where prices adjust within milliseconds, offer a natural laboratory for studying identification under extreme equilibrium feedback (Section 2.1), and closer collaboration between microstructure researchers and the causal-inference and dependence-modeling studies reviewed here would benefit both strands.

Finally, communication and transparency deserve greater attention. Clearly reporting identification assumptions, dependence patterns, multiple-testing corrections, and robustness checks would make financial research easier to understand and replicate. As this research increasingly informs policy, regulation, and risk management, overconfidence in uncertain results can be costly, making transparent treatment of uncertainty especially important.

8. Synthesis: A Framework for Applied Research

Table 1 places this survey against selected prior reviews in financial econometrics, summarizing how each treats identification, dependence, and uncertainty.

Building on the three-pillar framework and the interactions described in Section 5, the following four questions can help researchers and reviewers assess the credibility of an empirical finding in finance: (i) Identification: what specific assumption identifies the parameter of interest, and how would the conclusion change if that assumption were relaxed to a defensible set rather than a point? (ii) Dependence: what is the plausible effective sample size, $n_{\mathrm{eff}}$, once cross-sectional and temporal dependence are accounted for, and could the identifying variation itself reflect a common shock? (iii) Uncertainty: beyond the reported standard error, how sensitive is the result to reasonable alternative specifications, and how many such specifications were, in practice, considered before this one was reported? (iv) Convergence: does the finding hold up across independent datasets, sample periods, and research designs, or does it rest on a single specification? For policymakers and risk managers relying on this research, the same four questions apply directly: an identification strategy, a dependence structure, and an uncertainty quantification should each be stated explicitly enough that a result’s fragility, not only its point estimate, can be assessed before it informs a decision.

9. Conclusions

In this survey, we have organized financial econometrics around three issues that are fundamental to empirical work: identification, dependence, and uncertainty. Estimation techniques have substantially expanded the methodological tools available to empirical researchers. But better estimators do not, by themselves, make empirical conclusions more credible. That credibility ultimately depends on the assumptions behind the analysis and on how well those assumptions fit the setting being studied.

Financial data on asset prices, corporate decisions, and investor behavior are jointly determined in equilibrium, in interconnected markets, and are influenced by common shocks, institutional changes, and evolving expectations. As a result, empirical observations are rarely independent, identification is often partial or fragile, and uncertainty extends well beyond conventional sampling variability. In such settings, precise estimates can coexist with ambiguous economic interpretations, and as Section 5 makes explicit, the three pillars compound one another rather than operating in isolation.

Recognizing these constraints helps clarify when econometric methods can provide reliable insights. Identification remains central because estimates are meaningful only when the underlying assumptions are credible. Dependence affects how much independent information is actually contained in financial data and influences both inference and risk measurement. Uncertainty remains unavoidable because empirical results depend on the data, the model, and the assumptions used in the analysis.

Overall, credible empirical finance requires more than sophisticated statistical tools. Researchers must pay attention to identification assumptions, dependence structures, and the uncertainty surrounding empirical conclusions. The comparison in Table 1 and the accompanying checklist are offered as a starting point for making these considerations routine rather than exceptional in applied work. A stronger focus on these issues can improve how researchers interpret evidence on asset prices, corporate decisions, and financial stability. Progress in the field will depend not just on new methods, but on clearer thinking about what financial data can, and cannot, tell us about complex economic systems.