Tracking the Unseen: AI-Driven Dashboards for Real-Time Detection of Calendar Anomalies in Cryptocurrency Markets

Abstract

This study introduces a novel AI-powered Business Intelligence Dashboard System (AIBIDS) designed to detect and visualize calendar-based anomalies in cryptocurrency returns. Focusing on Bitcoin as a case study, the system integrates unsupervised machine learning algorithms to identify periods of abnormal market behavior across multiple temporal resolutions. The proposed system leverages a star-schema OLAP data warehouse, enabling real-time anomaly detection, dynamic visualization, and drill-down exploration of market irregularities. Empirical results confirm the presence of pronounced calendar effects in Bitcoin returns, such as heightened anomalies during Q1 and Q4, and reveal model-specific sensitivities to local versus global volatility. Our novel platform offers a practical, scalable innovation for investors, analysts, and regulators seeking to monitor cryptocurrency markets more effectively, and contributes to the emerging FinTech literature on AI-driven anomaly detection and behavioral market dynamics.

1. Introduction

The rapid rise in cryptocurrency markets, led by Bitcoin, has attracted increasing attention from investors, researchers, and regulators due to their high volatility, speculative nature, and unique behavioral patterns (Anastasiou et al., 2021; Bouri et al., 2022). Among these characteristics is their susceptibility to calendar-based anomalies—such as day-of-the-week or month-of-the-year effects—which have long been documented in traditional financial markets and are now emerging in the cryptocurrency domain (Baur et al., 2019; Naz et al., 2023; Qadan et al., 2022). Recent evidence suggests that cryptocurrencies exhibit systematic, non-random patterns across calendar intervals, implying persistent market inefficiencies, behavioral biases, or underlying structural trading dynamics (Barkai et al., 2024; Franco & Laurini, 2025).

Despite this empirical progress, practical tools for real-time detection, visualization, and interpretation of such calendar anomalies remain scarce. Most existing studies rely on static, backward-looking statistical techniques, which are inadequate for high-frequency, fast-moving markets like Bitcoin. As cryptocurrencies trade continuously across global time zones, they exhibit extreme price swings driven by sentiment, social media, and unexpected news events (Corbet et al., 2020). Furthermore, recent research confirm that return volatility varies significantly within the trading day (Franco & Laurini, 2025; Katsiampa et al., 2019), and that intraday inefficiencies challenge the Efficient Market Hypothesis (Algieri et al., 2025; Chu et al., 2019). These insights highlight the limitations of conventional models, and underscore the need for dynamic, automated systems capable of capturing nonlinear, high-frequency patterns and emerging anomalies as they emerge.

Motivated by these challenges, we develop a novel cloud-based, open-source AI-powered Business Intelligence Dashboard System (AIBIDS) to detect and monitor calendar anomalies in cryptocurrency markets. To demonstrate its capabilities, we apply AIBIDS to Bitcoin, a leading digital asset with well-documented anomaly patterns (Aharon & Qadan, 2019; Kinateder & Papavassiliou, 2021; Ma & Tanizaki, 2019; Naz et al., 2023). AIBIDS integrates four unsupervised machine learning and AI unsupervised algorithms—Isolation Forest (ISOF), Local Outlier Factor (LOF), One-Class SVM (OCSVM), and Long Short-Term Memory Autoencoder (AELSTM)—within a real-time, calendar-sensitive OLAP framework. The system continuously ingests updated price data via public APIs, detects anomalies across multiple time resolutions (daily, weekly, monthly, quarterly), and presents results through interactive drill-down dashboards. This approach offers users timely insights into evolving market behaviors that traditional models may fail to detect.

AIBIDS have several advantages and contributes directly to the literature on financial anomaly detection, market efficiency, and FinTech applications. First, the platform integrates state-of-the-art unsupervised AI models capable of identifying a wide range of anomaly types, from abrupt regime shifts to subtle, localized deviations. Second, it employs a multidimensional business intelligence architecture, allowing anomaly exploration across time layers such as year, quarter, month, and weekday—an approach that remains underexplored in crypto anomaly research. Third, all detection and aggregation operations are computed in-memory and in real time, enabling fast, interactive analysis without requiring complex backend processing. Finally, the system is fully automated, API-driven, open-source, and freely available online. It supports real-time updates, easy replication, and seamless integration with multiple data providers, making it a scalable tool for researchers, analysts, regulators, and FinTech practitioners. To the best of our knowledge, this is the first system to provide multidimensional calendar anomaly analysis using AI-based models in an open-access, user-friendly dashboard environment.

Beyond its practical implementation, this study advances the literature on financial anomalies by introducing artificial intelligence as a dynamic framework for detecting time-varying market inefficiencies. While prior research has documented various calendar anomalies in cryptocurrency markets (Veloso et al., 2025), these patterns are often unstable, nonlinear, and subject to structural shifts, making traditional econometric models insufficient. Our AI-based approach overcomes this limitation by providing an adaptive and scalable tool that continuously detects evolving anomalies, enabling investors to monitor irregular market behavior and enhance decision-making in real time. Overall, this framework deepens our understanding of cryptocurrency market inefficiencies in the presence of dynamic and shifting calendar effects.

The reminder of this paper is structured as follows: Section 2 outlines out materials and methods of proposed AI models architecture, BI dashboards system and experimental data. In this section, we also elaborate on the development and implementation of the BI system, offering a comprehensive overview of the practical implementation process, AI models training and evaluation results. Section 3 showcases the AIBIDS dashboard results and highlights anomaly calendars results for trained models. Section 4 summarizes main research findings and discusses the overall performance of implemented system. Finally, Section 5 concludes main research findings and outlines potential future research.

2. Materials and Methods

Market price anomalies can arise when investors overestimate (or underestimate) returns and underestimate (overestimate) risk. There is a growing interest in identifying calendar anomalies in cryptocurrencies, which are inherently more speculative and exhibit distinct characteristics. Given that both data scientists and financial analysts endorse the proposed methodology to analyze anomalies in financial markets. The main proposed methodology purpose is to formulate a multidimensional business intelligence dashboard visualization system AIBIDS that enables anomaly exploration across multiple temporal layers such as year, quarter, month, and weekday resolutions using state-of-the-art unsupervised AI anomaly detection algorithms.

2.1. AI Models Architecture

To detect anomalies in cryptocurrency markets, we trained three unsupervised machine learning models: Isolation Forest (ISOF), Local Outlier Factor (LOF), and One-Class Support Vector Machine (OCSVM). Unsupervised learning was employed because the collected Bitcoin dataset lacked labeled anomalies, making supervised approaches infeasible. ISOF, LOF, OCSVM, and an LSTM autoencoder (AELSTM) models were chosen for their complementary strengths: OCSVM captures deviations from learned cluster boundaries, LOF detects locally sparse outliers relative to their neighbors, ISOF isolates anomalous points through tree-based partitioning, and the LSTM autoencoder reconstructs temporal patterns in time series using recurrent neural networks with long short-term memory mechanisms.

Although Yahia et al. (2025) propose ISOF recommend ISOF as a baseline for anomaly detection in similar contexts, relying solely on a single model may risk overlooking nonlinear, time-dependent, or context-specific irregularities. Our framework therefore extends beyond ISOF by incorporating AELSTM and additional unsupervised models to provide broader methodological flexibility and more robust detection capabilities. The first three models are widely used in anomaly detection research for financial and high dimensional time series data, offering a strong balance between effectiveness and computational efficiency. The LSTM autoencoder provides a promising AI approach that identifies anomalies through reconstruction error based on learned temporal dependencies.

The input variable consisted of Bitcoin closing prices, and time series log return preprocessing was applied. Feature engineering included removing missing values and performing Min–Max scaling. The optimal parameter rates were determined using scikit-learn grid search optimization routine. A detailed explanation of the model hyperparameters is provided in Table 1, Table 2, Table 3 and Table 4.

Table 1. Isolation Forest ISOF model parameters.

Parameter	Value	Explanation
n_estimators	200–300	Increasing the number of trees generally improves model robustness and anomaly detection, especially for complex datasets. However, computational cost increases linearly with the number of trees. A n_estimators value 200 provides an effective balance between accuracy and efficiency.
max_samples	0.8	Specifies that 80% of the dataset is drawn to train each tree estimator. This ensures a representative sample for each tree while maintaining computational efficiency.
contamination	0.05–0.15	Defines the expected proportion of outliers in the dataset. Lower values indicate fewer anomalies.
max_features	log2	Determines the number of features to consider when splitting a node in each Isolation Tree. The max_features log2 value improve efficiency and reduce overfitting for smaller or moderately sized feature sets.

Table 2. Local Outlier Factor LOF model parameters.

Parameter	Value	Explanation
k	20–40	Specifies the number of neighbors to consider when computing the local outlier factor. A value of 20 is commonly used—it is large enough to capture local structure while small enough to preserve local patterns.
metric	Euclidean	Specifies the distance metric used to compute distances between data points. Euclidean distance is a common choice and generally performs well and computationally efficient when dealing with time series sparse data.
contamination	0.05–0.15	Anomalies are determined using the p-percentile of the Local Outlier Factor scores, corresponding to a contamination rate of 1–p.

Table 3. One Class Support Vector Machine OCSVM model parameters.

Parameter	Value	Explanation
nu	0.1	Represents the expected fraction of outliers in the data, implying that approximately 10% of data points are considered anomalies.
gamma	0.1	Kernel coefficient that controls the flexibility of the decision boundary. A smaller value, such as 0.1, produces a smoother and more generalized boundary.
kernel	rbf	The RBF kernel is widely used for nonlinear classification, offering a balance between flexibility and computational efficiency.
max_iter	−1	Sets the maximum number of iterations for optimization. A value of −1 allows the model to run until convergence.

Table 4. LSTM Autoencoder AELSTM model parameters.

Parameter	Value	Justification
latent_dim	4	The latent dimension controls the size of the compressed representation (i.e., the “bottleneck”). A small latent size like 4 forces the model to learn the most salient temporal patterns in the data, acting as a form of regularization and helping avoid overfitting.
loss	mae	Mean Absolute Error penalizes all errors linearly, making it robust to outliers and sharp reconstruction gaps. It is commonly used for reconstruction-based anomaly detection, where the absolute difference between original and reconstructed values is meaningful.
activation	relu	Rectified Linear Unit ReLU is a standard nonlinearity that introduces sparsity and helps mitigate the vanishing gradient problem.
optimizer	adam	Adam optimizer provides adaptive learning rates and much faster convergence, making it ideal for training recurrent models like LSTMs. It also adjusts learning rates per weight, which is helpful in deep autoencoders where different layers may learn at different rates.

Our methodology integrates these complementary algorithms to enhance anomaly detection, leveraging both classical unsupervised techniques and a deep learning-based generative model, providing a robust framework for uncovering unusual patterns in complex financial data.

Table 1, Table 2, Table 3 and Table 4 summarize the parameter settings and corresponding justifications for the ISOF, LOF, OCSVM and AELSTM models integrated into the AIBIDS dashboard system. These models were implemented using the open-source scikit-learn machine learning library in Python 3.13.

The Isolation Forest (Lesouple et al., 2021) algorithm identifies isolates observations through random feature and split selection. Each Isolation Tree partitions the data by randomly selecting a feature and a split value, with the ensemble of isolation trees producing an anomaly score for each data point. This approach is particularly effective for high-dimensional datasets and complements the other unsupervised models in capturing subtle deviations. Table 1 below outlines the parameter settings and corresponding justifications for the ISOF model implemented in the proposed system.

These considerations motivated the development of AIBIDS, which allows analysts to choose the most suitable anomaly detection model for a given dataset and temporal resolution. The system also supports comparative BI analysis and provides seasonal slicing with drill-down and drill-up capabilities across multiple temporal resolutions.

The Local Outlier Factor (LOF) method (Adesh et al., 2024) is an unsupervised machine learning algorithm used for anomaly detection. Its core principle is to identify anomalous data points by measuring the local deviation of each point relative to its neighbors. The algorithm assigns a local outlier factor score to every data point, quantifying the degree to which it can be considered as anomaly point. Table 2 below outlines the parameter settings and corresponding justifications for the LOF model implemented in the proposed system.

The One-Class Support Vector Machine (OCSVM) algorithm is an unsupervised machine learning method for anomaly detection that seeks to find abnormal points within a defined boundary. It classifies data into two groups using a decision function: the positive group represents normal points, while the negative group represents abnormal or anomaly points. Table 3 below outlines the parameter settings and corresponding justifications for the OCSVM model implemented in the AIBIDS dashboard system.

The Long Short Time Series LSTM time series autoencoder (Dip Das et al., 2024) AELSTM is a Generative AI algorithm that aims to learn an efficient, compressed representation of sequential data by encoding the input sequence into a lower-dimensional space and then reconstructing it as accurately as possible. It is particularly useful in anomaly detection, as sequences that cannot be accurately reconstructed may indicate anomalous behavior. Table 4 below outlines the parameter settings and corresponding justifications for the AELSTM model implemented in the AIBIDS dashboards system.

Table 5 outlines the architecture of the AELSTM model as implemented in the AIBIDS dashboard system for anomaly detection.

Table 5. LSTM Autoencoder AELSTM model architecture.

Layer	Output Shape	Notes
InputLayer	(None, 1, 1)	Time series input
LSTM	(None, 1, 16)	First encoder LSTM with 16 units
LSTM	(None, 4)	Encodes the input into 4 latent features
RepeatVector	(None, 1, 4)	Prepares latent vector for decoding
LSTM	(None, 1, 4)	First decoder LSTM
LSTM	(None, 1, 16)	Second decoder LSTM
TimeDistributed	(None, 1, 1)	Time series output reconstruction

The proposed AELSTM model is specifically designed for anomaly detection in univariate time series data. It encodes temporal patterns through stacked LSTM layers into a compact four-dimensional latent space and reconstructs the input using a symmetric decoder architecture. Anomalies are detected by measuring the reconstruction error, based on the assumption that normal patterns are accurately reconstructed, while anomalies result in higher error values.

As an unsupervised deep learning model, it belongs to the broader class of generative AI, where the decoder functions as the generative component that reconstructs input data from compressed representations. This approach leverages the model’s ability to learn the underlying structure of normal sequences without requiring labeled anomaly data, making it highly applicable to real-world anomaly detection tasks.

2.2. BI System Architecture

The proposed system was developed using a multidimensional star schema architecture. This architecture contains a fact table and calendar and cryptocurrency dimension tables.

The fact table stores information such as the cryptocurrency asset name, date, AI model, year, quarter, month, and weekday—each represented as a foreign key linked to its respective dimension table. The combination of these foreign keys forms the unique composite primary key of the fact table. Cryptocurrency closing prices were imported via the CoinGecko crypto data provider API.

The dimension tables include Asset, AI Model, Year, Quarter, Month, and Weekday. The time calendar dimension hierarchy is defined as Year/Quarter/Month/Weekday and Date dimensions. It enables drill-down and roll-up aggregations of anomaly measures and supporting multidimensional analysis of data anomalies.

The Return measure in the fact table was calculated using Equation (1), based on the corresponding Bitcoin closing prices for days t − 1 and t:

$$Return \left(t\right)= ln{\displaystyle \frac{Close \left(t\right)}{Close (t-1)}}$$

(1)

In the fact table, the Anomaly Model Indicator is a binary value where 1 indicates that an anomaly was detected by the predicted AI model, and 0 indicates no anomaly. The Anomaly High measure is a binary indicator that equals 1 when an anomaly is detected (value = 1) and the corresponding daily return is positive; otherwise, it equals 0. Similarly, the Anomaly Low indicator equals 1 when an anomaly is detected and the corresponding daily return is negative; otherwise, it is 0.

3. Results

This study utilized Bitcoin closing price time series data sourced and collected from the cryptocurrency data aggregator Coingecko.com. The collected data spans a substantial period from 1 January 2013, to 1 June 2025, comprising 4430 daily observations. The data was partitioned into training (80%) and testing (20%) subsets to support robust model fitting and evaluation. Figure 1 shows the AIBIDS architecture we use for anomaly detection in Bitcoin returns.

Figure 1. Cryptocurrencies AIBIDS star schema architecture.

Bitcoin exhibited a maximum daily return of 28.71% and a minimum of −43.37%. The median return of 0.15% suggests that Bitcoin generally offers modest but more consistent positive returns, reflecting increasing market stability over time.

Figure 2 presents the return dashboard for Bitcoin prices. The dashboard illustrates the high volatility of the Bitcoin market, with frequent spikes indicating abrupt price changes. Additionally, it displays descriptive statistics such as mean, maximum, minimum, standard deviation, and median returns.

Figure 2. Bitcoin returns dashboard with descriptive statistics.

3.1. AI Models Implementation

The proposed AI algorithms were implemented and evaluated using the Scikit-learn and Keras libraries. Model performance was validated using an 80/20 train–test split to ensure consistency and comparability across all anomaly detection approaches. The AELSTM model was trained and validated over 30 epochs with a batch size of 64, balancing sufficient learning iterations with computational efficiency. A validation split of 0.2 was applied, reserving 20% of the training data for evaluating model performance during training. The verbose = 1 parameter enabled progress updates at each epoch, providing real-time feedback on training and validation metrics.

Figure 3 illustrates that both training and validation loss for the AELSTM model decrease sharply during the initial epochs and gradually converge to near-zero values by approximately epoch 25. This pattern suggests that the model fits the data well and maintains strong generalization performance without signs of overfitting.

Figure 3. Training and validation loss curves for the AELSTM model. The graph shows a rapid decline in loss during early epochs, followed by gradual convergence, indicating effective learning and minimal overfitting.

In this study, the Bitcoin dataset did not contain predefined anomaly labels. Selecting a single “baseline” model is challenging, as no single method captures all aspects of anomalous behavior in volatile data. To address this, each model was first applied to the full dataset, and the anomalies it detected were treated as its baseline (interpolated) anomalies. We then trained each model on 80% of the data and predicted anomalies on the remaining 20%, producing extrapolated outputs. Thus, for each model we obtained two anomaly sets for the test portion: interpolated and extrapolated. Precision, Recall, and F1 metrics were calculated between these outputs, as classification metrics are better suited for evaluating unsupervised anomaly detection models. This procedure motivated our use of a BI system, which enables a clearer and more comprehensive examination of anomalous behavior across models. In our view, this approach provides a more meaningful assessment of each model’s anomaly detection capability.

Table 6 presents the evaluation metrics for the fitted AI anomaly detection models applied to Bitcoin returns. An evaluation based on the precision metric further confirms the superiority of the OCSVM model, which consistently attains the highest precision value of 1, indicating strong performance in accurately identifying true anomalies. Conversely, an evaluation focusing on the recall metric shows that the AELSTM model consistently outperforms the others, demonstrating greater effectiveness in detecting true anomalies while minimizing false negatives. The AELSTM and OCSVM models consistently outperform the ISOF and LOF models, achieving the highest F1 values 0.85 and 0.81.

Table 6. Evaluation Metrics of Anomaly Detection Models on the Bitcoin Test Set. This table summarizes the performance of four AI models—ISOF, LOF, OCSVM, and AELSTM—on the Bitcoin test set using Precision, Recall and F1 metrics.

Model	Test Set	Precision	Recall	F1
ISOF	886	0.77	0.64	0.7
LOF		0.41	0.59	0.49
OCSVM		1	0.68	0.81
AELSTM		0.89	0.82	0.85

These findings suggest that OCSVM and AELSTM are the most effective models for detecting anomalies in Bitcoin data, positioning them as the preferred default models for future implementations of the AIBIDS.

3.2. AI Models Evaluation

The evaluation reveals distinct anomaly detection characteristics for each AI model. Each model focuses on different aspects of global data deviations and identifies numerous anomalies, particularly during periods of high volatility and market spikes, as presented in Figure 4.

Figure 4. Bitcoin anomalies detected by AIBIDS system dashboards (Daily Returns). Notes: This figure illustrates daily Bitcoin return anomalies identified by AI-BIDS models (AELSTM, OCSVM, ISOF, LOF). Black lines mark known anomaly dates; red lines and dots denote detected anomalies and associated return values.

The AIBIDS dashboards in Figure 4 illustrate the performance of the AELSTM, OCSVM, ISOF, and LOF models in detecting anomalies in Bitcoin daily returns. The dashboards clearly display the dispersion of red anomaly points relative to normal data. Each model exhibits distinct anomaly detection patterns. The AELSTM model identifies more clustered anomalies and captures a wider spectrum of fluctuations throughout the time series, demonstrating sensitivity to both large and small variations in returns. Conversely, the OCSVM model detects anomalies across a broader range of time periods, including relatively stable intervals, suggesting higher sensitivity to dispersed market fluctuations. The ISOF model is more effective at capturing large, global anomalies and detects more anomalies during significant market events, while the LOF model identifies fewer anomalies overall and focuses primarily on local deviations.

3.3. AIBIDS Calendar Dashboards

All models, to varying degrees, show a cyclical pattern in anomaly scores, generally peaking and then declining. This suggests that periods of higher anomaly are followed by periods of lower anomaly, which aligns with the cyclical nature often observed in cryptocurrency markets (e.g., bull and bear cycles).

The overall trends across all models indicate a general decrease in anomalies over years. Peaks of anomaly across the years in Figure 5 may be explained as follows:

Figure 5. Bitcoin anomalies AIBIDS dashboard annual layer returns using AELSTM, OCSVM, ISOF and LOF models.

2013: ISOF, OCSVM and AELSTM acronymized as ICA models show a high anomaly score. This observation corresponds with Bitcoin’s significant price surge and subsequent crash in 2013, which would certainly be considered anomalous compared to its prior existence.

2017: A very pronounced peak is observed across ICA models. This strongly correlates with the massive Bitcoin bull run (crypto rush) of 2017, where its price soared to nearly $20,000, creating numerous anomalous price movements and market behaviors.

2019: COVID-19 pandemic in early 2020 also coincides with significant anomalies detected by ICA models, reflecting the heightened volatility and shifts in investor behavior as global markets reacted to the crisis. During this period, cryptocurrencies experienced dramatic price swings, as investors grappled with uncertainty, leading to increased trading volumes and market instability.

2021: A pronounced peak is observed for ICA models, coinciding with the bull market during which Bitcoin reached new all-time highs (above $60,000), resulting in notable anomalous activity.

2022: The escalation of the Russia–Ukraine conflict corresponds with detectable market irregularities, likely driven by increased reliance on digital assets amid sanctions and financial instability.

2025: The ICA models show an upward trend in anomaly scores from 2023 into 2024, followed by a projected decline into 2025. This could suggest anticipation of increasing anomalous activity leading up to or during a potential bull market in 2024, then tapering off.

In contrast, Local Outlier Factor LOF model behaves quite differently from the other three models. It generally reports lower anomaly scores, particularly during the major bull peaks (2013, 2017, 2021) where the other models spike significantly. Its peaks are more subdued, and its troughs are less pronounced. It shows a relatively higher anomaly score in 2014–2015 and 2019 compared to the other models during those periods. This difference suggests that LOF might be detecting different types of anomalies or is less sensitive to the large-scale price movements that the other models pick up. It might be focusing more on “local” deviations within specific time windows rather than overall market cycles. For instance, it might identify subtle but unusual trading patterns that occur even during relatively stable periods, or it might be less influenced by the magnitude of price changes and more by the distinctiveness of data points in local neighborhoods.

The chart in Figure 6 shows that the ICA models track each other very closely, exhibiting almost identical patterns. They start Q1 with the highest anomaly scores (around 125–133), dip sharply to their lowest point in Q3 (around 75–85), and then rebound strongly in Q4, reaching levels close to or even exceeding Q1. This strong correlation suggests these models are sensitive to similar types of quarterly market dynamics. Given Bitcoin’s history, Q4 is often a period of significant price movements and sometimes the start or continuation of bull runs, which would naturally generate higher anomaly scores. Peaks of anomaly across the years may be explained as follows:

Figure 6. Bitcoin anomalies AIBIDS dashboard quarters layer returns using AELSTM, OCSVM, ISOF and LOF models.

Q1: High anomaly, possibly due to post-holiday trading, new year capital flows, or the continuation of trends from the previous year’s Q4.

Q2–Q3: Lower anomaly, potentially representing a period of consolidation, lower volatility, or “summer doldrums” in traditional markets that extend to crypto.

Q4: High anomaly, often a period historically associated with major Bitcoin price movements, increased speculative activity, and the run-up to year-end.

In conclusion, this chart provides valuable insights into the average quarterly distribution of Bitcoin anomalies as detected by different algorithms. It clearly shows a tendency for anomalies to be higher in Q1 and Q4 and lower in Q2 and Q3, aligning with some historical market observations. The distinct behavior of the LOF model further emphasizes the nuanced nature of anomaly detection.

The chart in Figure 7 strongly supports the presence of a monthly pattern in Bitcoin’s anomalous behavior. The mid-year (summer) months appear to be typically calmer, while the end of the year is more active and anomalous.

Figure 7. Bitcoin anomalies AIBIDS dashboard monthly layer returns using AELSTM, OCSVM, ISOF and LOF models.

The peaks in May and especially November-December could correlate with historical periods of increased market volatility, price discovery, and significant news events. These months have often seen major Bitcoin price movements.

The consistent dip from June to October across most models suggests that the cryptocurrency market, like traditional markets, might experience periods of reduced activity and fewer “surprising” events during summer and early autumn. The consistent alignment of ICA models reinforces their sensitivity to broad market trends and significant deviations.

The LOF model, by showing a flatter and sometimes differently timed pattern, offers a more nuanced view, potentially detecting subtle anomalies that are not related to large-scale market swings. It might be detecting anomalous local movements or patterns within otherwise “normal” periods.

Finally, if these patterns hold true over time, such a chart could inform trading strategies, risk management, or simply provide context for expected market behavior throughout the year. For instance, expecting higher anomalies in Q4 and lower in Q3 might influence when one looks for specific trading opportunities or prepares for increased volatility.

The chart in Figure 8 presents that Bitcoin anomaly levels are not uniform throughout the week. Traders or analysts might prepare for different levels of “unexpected” activity depending on the day of the week. For example, Mondays might be less anomalous, while mid-weekdays like Tuesday and Friday could be more active. ICA models represent Tuesday as anomaly prominent peak. This could reflect the full swing of trading activity resuming after the weekend and Monday’s relative calm. ICA models notably peaks on Friday, which can be a volatile day in financial markets as positions are closed or new ones opened before the weekend. The pattern for Saturday is generally a decline for most models, suggesting that weekends, while potentially having some distinct trading patterns, might have overall lower “anomalousness” compared to the peak weekday activity. Similarly, to previous higher time resolutions, LOF model’s consistently lower and smoother curve indicates it is either less sensitive to the types of large, rapid movements that the other models identify as anomalous, or it is detecting more subtle, local deviations from typical patterns that are not tied to overall Bitcoin market volatility.

Figure 8. Bitcoin anomalies AIBIDS dashboard weekdays layer returns using AELSTM, OCSVM, ISOF and LOF models.

4. Discussion

The proposed AIBIDS dashboard system enhances the understanding of Bitcoin’s anomalous behavior by revealing distinct patterns across various calendar resolutions. The monthly anomaly chart offers a granular view of Bitcoin’s cyclical behavior, highlighting relatively calmer mid-year periods and more active, anomaly prone year-end periods. These trends broadly align with observed market seasonality. Additionally, mid-weekdays often exhibit higher anomaly levels, while Mondays and Saturdays tend to be relatively calmer across several models.

Overall, the trends observed across all models indicate a general decrease in anomalies over the years, suggesting improved market stability or a reduction in irregular behavior within the underlying data. Similar findings were reported by Veloso et al. (2025) supporting the notion that Bitcoin is transitioning from a speculative asset to a more stable financial instrument. Future research could investigate whether similar seasonal and structural patterns are present in other cryptocurrencies, offering a broader understanding of digital assets’ roles in the evolving financial ecosystem.

These findings support the consistency and reliability of the implemented anomaly detection models within the AIBIDS dashboard system. Using this system, users can identify deeper anomalous patterns in the data across different temporal resolutions and leverage various statistical indicators. For example, the user can determine whether a detected anomaly exhibits seasonal behavior or, alternatively, represents a one-time temporal spike. The system also enables assessment of how other models influence or characterize a specific anomaly and the extent to which that anomaly is significant. From a practical standpoint, incorporating new data sources or models into the system is straightforward. Consequently, the proposed system offers several advantages, as it allows financial analysts to detect, examine, and interpret anomalous data using advanced business-intelligence analytical operations such as drill-down, drill-up, and related techniques.

5. Conclusions

In conclusion, the proposed AIBIDS dashboard system effectively visualizes the presence and intensity of anomaly patterns in Bitcoin over time, as detected by three AI models. The strong correlation between high anomaly scores and Bitcoin’s major bull markets (2013, 2017, 2021) suggests that the One-Class Support Vector Machine, Isolation Forest, and LSTM Autoencoder models primarily identify periods of extreme price movements and heightened market excitement as anomalous.

The observed reduction in anomaly frequency over the years indicates increasing market maturity and potentially enhanced systemic resilience. This trend provides valuable insights for investors seeking to navigate short-term opportunities and for regulators aiming to understand and respond to evolving market dynamics. Future research could extend this analysis to other cryptocurrencies, offering a broader perspective on the role of digital assets in the modern financial ecosystem. Such investigations would further clarify Bitcoin’s evolving position within the global economic landscape.

While anomalies identified by the models at the original daily resolution do not exhibit clear cyclical seasonality, aggregating these anomalies into higher temporal resolutions—such as day-of-week or monthly intervals—reveals distinct seasonal patterns. This finding highlights the significant influence of time resolution on the observed behavior of Bitcoin prices. Accordingly, the development of AIBIDS that incorporate automatic hierarchical aggregation across multiple temporal resolutions holds both theoretical and practical value. Such systems can equip data scientists and financial analysts with advanced analytical tools, enabling deeper insights into temporal structures and market dynamics.

6. Limitation and Future Work

The limitations of this study indicate that the proposed AIBIDS currently includes only Bitcoin cryptocurrency data at a daily resolution and does not contain ground-truth anomaly labeling. From a practical perspective, it is possible to extend the system by adding other cryptocurrencies and additional temporal resolutions. As noted earlier, due to the lack of ground-truth data, imputation was performed using an inverse interpolation model. Sensitivity checks of several model parameters were also conducted using scikit-learn grid search; however, more advanced optimization techniques could be applied to further enhance the evaluated models.

Future research may examine how specific models detect anomalies within financial context. This will require additional text-based financial data and the development of multimodal Generative AI models to incorporate such contextual information.