Market Regimes#
Financial markets do not behave uniformly across time. They cycle through distinct regimes — bull and bear markets, high and low volatility periods, trending and mean-reverting environments, expansion and contraction phases — with each regime exhibiting different return distributions, correlations, and factor sensitivities. Regime detection aims to identify the current market state in real time, enabling adaptive trading systems that adjust strategy parameters, position sizes, or asset allocation based on the inferred regime.
The research here covers the foundational Markov regime-switching framework introduced by Hamilton (1989), regime-switching ARCH/GARCH models for volatility, modern machine learning approaches to regime identification (Wasserstein clustering, jump models, HMM), and practical applications to Bitcoin, VIX term structures, and multi-asset portfolios. A recurring theme is the challenge of real-time regime detection — the tension between using only past data (filtered probabilities) and the inevitable lag in regime identification that limits practical utility.
Related topics include Trend Following for regime-conditional trend strategies, Momentum for momentum timing with regime filters, and Volatility Modeling for volatility-based regime characterization.
A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle#
James D. Hamilton introduces the Markov-switching autoregressive model, where the parameters of an autoregression are governed by a discrete-state Markov process whose realizations are not directly observed. The framework allows the mean growth rate of a nonstationary series to shift between states, with the timing and duration of regimes inferred from the data via maximum likelihood. The paper applies the model to postwar U.S. real GNP growth, identifying distinct expansion and contraction regimes that align closely with NBER business cycle dates. Published in Econometrica, Vol. 57, No. 2, March 1989.
Our summary: this is the foundational paper for all regime-switching time-series models in economics and finance. Every subsequent application of Markov-switching to volatility, asset returns, or macro variables traces back to this framework. The key insight for quantitative finance is that a single linear model often masks fundamentally different dynamics across market states, and the Hamilton filter provides a principled, real-time way to estimate which regime the market is in at each point.
Autoregressive Conditional Heteroskedasticity and Changes in Regime#
James D. Hamilton and Raul Susmel extend the Markov regime-switching framework to ARCH volatility models, allowing the parameters governing conditional heteroskedasticity to shift between discrete states. The paper models weekly returns on U.S. stock indices, permitting the ARCH intercept and persistence parameters to change with an unobserved Markov state. Published in the Journal of Econometrics, Vol. 64, 1994.
Our summary: this paper is the direct bridge between Hamilton’s 1989 regime-switching framework and modern volatility modeling. By allowing ARCH parameters to switch between states, it formalizes the intuition that volatility clustering is not uniform over time but instead governed by latent market regimes. For realized-volatility forecasting and crypto volatility work, the Hamilton-Susmel model is the ancestor of all Markov-switching GARCH variants.
Clustering Market Regimes Using the Wasserstein Distance#
This paper introduces Wasserstein k-means clustering for identifying market regimes from return distributions. Unlike standard clustering that uses Euclidean distance on summary statistics, the Wasserstein approach operates directly on distributional properties, capturing the full shape of the return distribution in each period and identifying regime changes that conventional methods miss.
Rethinking Trend Following: Optimal Regime-Dependent Allocation#
This paper develops a framework for regime-dependent time-series momentum allocation, showing that the optimal trend signal depends on the current market regime. The authors demonstrate that adapting trend parameters and allocations to the identified regime generates superior risk-adjusted performance compared to static trend following implementations.
Dynamic Mean-Variance Portfolio Allocation under Regime-Switching Jump-Diffusions with Absorbing Barriers#
This paper solves the dynamic mean-variance portfolio allocation problem under a regime-switching jump-diffusion model with absorbing barriers representing ruin. The optimal allocation policy depends on the current regime and adapts to regime transitions, providing a rigorous framework for regime-aware portfolio construction under realistic market dynamics.
Regularised jump models for regime identification and feature selection#
A regime modelling framework that jointly performs regime identification and feature selection using regularised jump models. The framework identifies the active regime from market and macroeconomic variables while simultaneously selecting features that best distinguish between regimes, improving both interpretability and regime identification accuracy compared to standard jump models.
Dynamic Asset Allocation with Asset-Specific Regime Forecasts#
This article introduces a novel hybrid regime identification-forecasting framework that leverages both unsupervised and supervised learning to generate tailored regime forecasts for individual assets. The statistical jump model derives regime labels, and a gradient-boosted decision tree classifier predicts future regimes using return features and cross-asset macro-features. The framework is applied to a multi-asset portfolio comprising twelve risky assets from 1991 to 2023.
Forecasting Market Regimes with the sUSDe Term Structure#
An exploration of how the sUSDe term structure on Pendle can serve as a forward-looking signal for crypto market sentiment and regime forecasting. The article explains how crypto prices are heavily influenced by leveraged trading in perpetual futures, and how Ethena’s sUSDe effectively captures the basis from funding rates through delta-neutral strategies.
The article demonstrates that the term spread—the difference between back month and front month implied yields—is highly correlated with underlying yield regimes and produces a stronger signal for returns than the underlying yield alone. Historical analysis supports using this term structure as a leading indicator for changes in cost of carry and BTC price levels.
By Luke Leasure.
Testing the January Barometer as a Regime Filter for Crypto Trend Models#
Research Article #46 from Trading Research Hub. Tests the January Barometer effect in cryptocurrency markets. Originally introduced by Yale Hirsch in 1972 for equities, the January Barometer suggests that January’s market performance predicts the trend for the rest of the year.
The article applies this calendar anomaly to crypto, testing whether January’s performance in Bitcoin and other cryptocurrencies has predictive power for the remaining eleven months. It also examines whether the effect can be incorporated as a regime filter into a systematic trend-following model.
By Pedma.
Adjusting Bitcoin Strategy Exposure Based on Volatility Regimes#
Research Article #26 from Trading Research Hub. Studies how volatility regimes affect trading strategy performance in Bitcoin. The study examines how adjusting portfolio exposure based on the current volatility environment can improve risk-adjusted returns.
The article tests systematic approaches to identifying and responding to different volatility regimes, from calm to highly volatile periods. By targeting specific volatility conditions or adjusting position sizes based on the prevailing regime, the strategy aims to reduce drawdowns during turbulent periods while maintaining exposure during favorable conditions.
By Pedma.
Systematic Regime Detection for Momentum Strategy Timing#
An article on the systematic approach to regime targeting for momentum trading strategies. The article argues that trading against the general market direction is one of the most common causes of losses, and presents a framework for identifying favorable market regimes.
The article covers how to define and detect market regimes, how to adjust strategy exposure based on the current regime, and how to avoid trading during unfavorable conditions. The practical framework helps momentum traders align their positions with the broader market trend while knowing when to reduce exposure.
By Pedma.
Detecting VIX Term Structure Regimes#
A practical guide to analyzing the VIX term structure using Principal Component Analysis (PCA) and Hidden Markov Models (HMM) in Python. The article argues that a single VIX number tells you almost nothing about where risk sits in the market, but the term structure reveals whether the crowd expects a crisis next week or three months from now.
The methodology uses PCA to reduce the dimensionality of the VIX term structure into interpretable components (level, slope, curvature), then applies a Gaussian Hidden Markov Model to classify the slope series into distinct regimes including contango, backwardation, and transition states. A complete Python notebook is provided for reproducing the analysis.
By Cristian Velasquez.
Mentioned by QuantSeeker in this discussion.
Regime Detection with HMM: Critical Analysis#
A critical examination of hidden Markov model-based regime detection for systematic trading, covering both the theoretical appeal and practical limitations of HMM approaches. The post discusses the strengths of the method — including its principled probabilistic framework for latent state estimation — and several critical concerns including state number sensitivity, real-time detection lag, and the challenge of distinguishing regime detection from regime prediction.
Several aspects warrant scrutiny: HMMs require the number of hidden states to be specified a priori; real-time regime detection lags regime transitions substantially; the 70% probability threshold for reducing leverage is presented without justification; and transaction costs from regime-switching strategies can erode returns significantly if regime changes are detected frequently or noisily.
By Bongani Mayaba.
Building Regime-Robust Trading Systems Through Market Cycle Analysis#
An article on building robust long-term trading systems, explaining the market dynamics of inventory accumulation and distribution. The article describes how a trading system with a reliable signal acts like a lighthouse, attracting market participants and creating the conditions for profitable trading through systematic inventory management.
The article emphasizes the importance of distinguishing between systems that work in specific market conditions (like 2020’s bull market) versus those that are genuinely robust across different regimes.
By Pedma.
Architecting Market Regime Clusters for Adaptive Portfolio Construction#
A practical guide to implementing market regime clustering as infrastructure for adaptive portfolio construction. The article argues that classical quantitative models fail because they assume stationary markets, and proposes unsupervised learning approaches to partition financial time series into distinct states with unique risk-return distributions. The recommended feature stack goes well beyond raw returns — rolling skewness and kurtosis, volatility cones, cross-asset dynamics (CDX vs VIX), and microstructure signatures (mid-price autocorrelation, bid-ask imbalance) — with manifold learning (t-SNE or autoencoders) for dimensionality reduction before clustering.
Three algorithmic approaches are compared: Gaussian Mixture Models for soft clustering that reflects gradual transitions, HDBSCAN for non-parametric density-based detection with transition states classified as noise, and Jump-Diffusion Hidden Markov Models for capturing instantaneous regime shifts. The article identifies three canonical regimes — Low-Vol Growth (suited for leveraged risk-parity), High-Vol Inflation (requiring minimum-variance pivots), and Tail Crash (triggering convexity overlays). A particularly sharp observation is that widely adopted “Risk-Off” signals can become self-fulfilling prophecies, ceasing to be leading indicators and instead becoming coordinating mechanisms for liquidation. The validation framework uses BIC optimisation, Viterbi path analysis, MST comparisons, and a “Regime Coherence Test” on realised volatility.
By Systan.
Regime changes in Bitcoin GARCH volatility dynamics#
This paper (Ardia, Bluteau & Ruede, Finance Research Letters, 2019) tests for the presence of regime changes in the GARCH volatility dynamics of Bitcoin log-returns using Markov-switching GARCH (MSGARCH) models, and compares them against traditional single-regime GARCH specifications for predicting one-day-ahead Value-at-Risk (VaR). Estimation is Bayesian, with the posterior used both to fit the models and to generate the VaR forecasts. The authors consider symmetric and asymmetric conditional-variance dynamics and Normal versus Student-t innovations across one- and two-state specifications, allowing volatility persistence and leverage effects to differ by regime.
Our summary: the paper provides clean statistical evidence that Bitcoin volatility is governed by distinct regimes rather than a single GARCH process. The best in-sample fit is a two-state model with asymmetric dynamics and Student-t innovations, and the analysis surfaces an inverted leverage effect (positive shocks raising volatility more than negative ones) present in both regimes. Most importantly for practitioners, MSGARCH models clearly outperform single-regime GARCH when forecasting VaR.
Key metrics: the central empirical claims are (i) strong evidence of regime changes in the GARCH process, (ii) a two-state asymmetric Student-t specification as the best in-sample model, (iii) an inverted leverage effect in both regimes, and (iv) MSGARCH dominating single-regime GARCH in out-of-sample one-day-ahead VaR forecasting, assessed by VaR backtesting. The study is on a single asset (Bitcoin) with daily data through the late-2010s. It reports volatility/VaR statistical performance rather than trading metrics (no Sharpe, no P&L); the authors’ MSGARCH framework is available through the R package MSGARCH, supporting reproduction.
Critical view: this is a credible, well-scoped paper from authors who maintain the relevant estimation software, and the headline result, regime-switching improving VaR forecasts, is evaluated genuinely out-of-sample via VaR backtesting, which is the right test. The main limitations are inherent to the setting: a single, short, highly volatile crypto series limits generalization, and the “best model” is chosen after comparing several GARCH variants, so some model-selection/data-snooping risk attaches to the specific winning specification (mitigated, but not eliminated, by the out-of-sample VaR focus). The inverted leverage effect is an interesting but sample-specific finding. There is no trading-strategy claim and no transaction-cost analysis, which is appropriate since the paper is about risk measurement, not returns. Overall this is one of the more methodologically sound entries: modest in its claims and validated where it matters.
Modelling the volatility of cryptocurrencies using Markov-Switching GARCH models#
This paper (Caporale & Zekokh, Research in International Business and Finance, 2019) conducts a large-scale model comparison to find the best specification(s) for the volatility of four major cryptocurrencies: Bitcoin, Ethereum, Ripple, and Litecoin. The authors fit more than 1,000 GARCH-type models, spanning single-regime and Markov-switching (regime-dependent) variants with different conditional-variance dynamics and innovation distributions, and generate one-step-ahead forecasts of Value-at-Risk (VaR) and Expected Shortfall (ES) for each. Model adequacy is assessed via VaR and ES backtesting together with the Model Confidence Set (MCS) procedure, which formally selects the statistically superior subset of models while controlling for multiple comparisons.
Our summary: the key message is that standard single-regime GARCH models are inadequate for crypto risk, and that allowing for regime switching materially improves tail-risk forecasts. Across all four cryptocurrencies, two-regime models produce better VaR and ES predictions than single-regime models. The use of the MCS procedure across 1,000+ models is the methodological strength, since it guards against picking a single lucky specification.
Key metrics: more than 1,000 GARCH models are estimated across four cryptocurrencies (BTC, ETH, XRP, LTC); evaluation is via one-step-ahead VaR and ES backtesting plus the Model Confidence Set. The headline finding is that two-regime (Markov-switching) GARCH specifications dominate single-regime ones in VaR/ES accuracy for all four assets. The paper reports risk-forecasting/statistical performance (VaR/ES backtests, MCS membership) rather than trading metrics; no Sharpe ratios or strategy P&L are given. The estimation leans on the MSGARCH R package, aiding reproducibility.
Critical view: this is a rigorous and appropriately cautious study. The combination of out-of-sample VaR/ES backtesting with the Model Confidence Set directly addresses the main pitfall of fitting hundreds of GARCH variants, namely data snooping and model-selection bias, because MCS is explicitly designed to control for searching over many models. Testing four cryptocurrencies rather than Bitcoin alone strengthens external validity relative to single-asset studies. Remaining caveats are modest: the crypto samples are short and highly non-stationary, so even out-of-sample windows cover an unusual, bubble-prone period; ES backtesting is statistically harder than VaR and those conclusions rest on relatively few tail observations; and the analysis is about risk measurement, so it makes no claim about tradability and ignores transaction costs (reasonably so). On balance, the regime-switching-improves-tail-risk conclusion is well supported and among the more reliable results in this set.
Modelling and Predicting the Conditional Variance of Bitcoin Daily Returns: Comparison of Markov Switching GARCH and SV Models#
This paper by Dennis Koch, Vahidin Jeleskovic and Zahid I. Younas compares two families of volatility models for Bitcoin daily log returns: Markov-Switching GARCH (MS-GARCH) models and stochastic volatility models, specifically the Stochastic Autoregressive Volatility (SARV) family. The authors adopt a two-stage estimation approach that separates the estimation of the mean-equation coefficients from those of the variance equation, and they evaluate the competing specifications through out-of-sample forecasting of Bitcoin’s conditional variance rather than relying solely on in-sample fit.
Our summary: the central finding is that stochastic volatility models, particularly SARV specifications, outperform MS-GARCH models in forecasting Bitcoin price volatility. A notable and somewhat deflationary result is that even simple single-regime GARCH models can outperform the more elaborate Markov-Switching GARCH variants for predicting the variance of Bitcoin log returns. The contribution is therefore more cautionary than triumphant: added regime-switching machinery does not automatically buy better volatility forecasts for crypto.
Key metrics: the paper compares MS-GARCH against SARV/SV using out-of-sample forecast loss criteria, but specific log-likelihood, AIC/BIC, MSE/QLIKE loss values, the exact sample period, and the number of regimes are not recoverable from the public abstract. The paper is a 2024 arXiv preprint; no trading or financial performance metrics (Sharpe, hit rate, VaR/ES backtests) are reported, as the study is purely about variance forecasting accuracy. No explicit data or code repository is mentioned.
Critical view: the headline conclusion that simple GARCH and SARV beat MS-GARCH is credible and aligns with a recurring finding in the crypto-volatility literature that regime-switching GARCH often overfits in-sample and degrades out-of-sample. Because the authors emphasize out-of-sample evaluation, the comparison is less prone to the in-sample overfitting that flatters regime models, which is a genuine strength. The main caveats are familiar to this genre: a single asset (Bitcoin) and likely a single contiguous daily sample limit external validity; the two-stage estimation may discard efficiency relative to joint estimation and can bias variance-equation inference if the mean is misspecified; and conclusions about model ranking can be sensitive to the chosen loss function (MSE versus QLIKE) and to whether a formal Model Confidence Set test is applied. Without the published metrics it is hard to judge whether the SARV advantage is statistically significant or merely a point-estimate ordering. As a directional result it looks real; as a precise quantitative claim it is under-documented in the accessible material.
A Quantile Spillover-Driven Markov Switching Model for Volatility Forecasting: Evidence from the Cryptocurrency Market#
This paper by Fangfang Zhu, Sicheng Fu and Xiangdong Liu (Mathematics, MDPI, 2025) proposes a regime-switching realized-volatility model that incorporates time-varying, quantile-based spillover effects. The authors first build a dynamic spillover factor by identifying, across different quantile levels, the most influential contributors to Bitcoin’s realized volatility, distilling a complex and time-varying network structure into a single economically interpretable endogenous driver. This quantile-layered spillover factor is then embedded into a HAR (Heterogeneous Autoregressive) realized-volatility model with a time-varying transition probability (TVTP) Markov-switching mechanism, yielding a TVTP-MS-HAR model in which spillovers both explain regime transitions and sharpen regime identification.
Our summary: the contribution is the quantile-driven spillover factor: instead of feeding a high-dimensional spillover network directly into the model, the authors compress it into a single interpretable, endogenous regime driver that captures heterogeneous spillover paths under different market conditions. Empirically the proposed TVTP-MS-HAR model is reported to deliver superior out-of-sample volatility forecasts relative to standard HAR-type benchmarks, while better identifying state-dependent spillovers and nonlinear dynamics in the cryptocurrency market.
Key metrics: the paper reports forecast-comparison results against standard HAR/MS-HAR benchmarks, but the publicly accessible abstract does not expose the concrete loss values; specific MSE/QLIKE/MAE numbers, Model Confidence Set outcomes, out-of-sample R-squared, the exact sample period, and the full cross-section of cryptocurrencies are not recoverable from the abstract alone. No trading/financial performance metrics (Sharpe, VaR/ES backtests, hit rate) are described; the evaluation is statistical forecast accuracy on realized volatility. Data are realized-volatility series centered on Bitcoin; code availability is not indicated.
Critical view: the framework is intellectually appealing and the TVTP mechanism is a principled way to make transition probabilities respond to an observable driver, which is more defensible than constant-probability MS models. The quantile-spillover compression also addresses a real problem (dimensionality of network inputs). That said, this design has many degrees of freedom: the choice of quantile levels, how the “most influential” contributors are selected, the HAR lag structure, and the TVTP link function all create scope for data-snooping and in-sample tuning, so a clean, pre-specified out-of-sample protocol and an MCS-style test across many models are essential to trust the “superior forecasting” claim. Constructing the spillover factor using quantile information risks look-ahead bias if any full-sample quantile estimation leaks into the forecasting window. Realized-volatility forecasting also rewards complexity in-sample, and HAR benchmarks are easy to beat marginally; whether the gains survive transaction-cost-aware use or out-of-sample across multiple coins and sub-periods is not answerable from the abstract. The idea is solid and publishable; the magnitude and robustness of the forecasting edge need the full results tables to verify.
Regime-Switching Behaviour in US Equity Indices: Two State Model With Kalman Filter Tracking and Finite State Machine Trading System#
This Master of Applied Science thesis by Timothy Little (Ryerson University, now Toronto Metropolitan University, 2012, Electrical and Computer Engineering) develops a time-varying two-state regime-switching model for US equity index daily returns. Rather than estimating fixed regime parameters, the model’s parameters are tracked recursively over time using a Kalman filter, treating the latent regime dynamics as a state-space signal-processing problem. The author reports that this recursive, time-varying formulation improves model fit relative to static approaches. The information extracted from the fitted model is then used to construct a Finite State Machine (FSM) trading system that switches between long/defensive states according to the inferred regime.
Our summary: the work bridges signal processing and quantitative finance by casting equity regime detection as Kalman-filter state tracking and turning the regime signal into a deterministic Finite State Machine trading rule. The headline backtest claim is dramatic: the FSM trading system is reported to outperform buy-and-hold by more than 15,000% on the Dow Jones Industrial Average over 1928-2012, with similar outperformance on the S&P 500 and NASDAQ Composite.
Key metrics: indices are the DJIA, S&P 500, and NASDAQ Composite; the sample is daily returns spanning 1928-2012 for the DJIA (shorter for the others). The model is a two-state regime-switching model with Kalman-filter parameter tracking. Reported performance is a cumulative backtested return exceeding 15,000% above buy-and-hold for the DJIA; the thesis emphasizes improved model fit but does not give Sharpe ratios, drawdowns, turnover, hit rate, or transaction-cost figures. DOI 10.32920/ryerson.14651955.v1.
Critical view: the 15,000%-over-buy-and-hold figure should be read with strong skepticism. Cumulative-return multiples over an 84-year window are extremely sensitive to compounding and to a handful of well-timed exits around the 1929, 1987, 2000 and 2008 crashes, so a single regime signal that sidesteps a few crashes can mechanically produce enormous headline outperformance without implying a robust, repeatable edge. The most serious concern is in-sample versus out-of-sample evaluation: if the regime model and FSM thresholds were fitted on the same 1928-2012 history they are backtested on, the result is curve-fit and look-ahead biased. A Kalman filter run as a smoother (using future data) versus a true real-time filter is the crucial distinction; only causal, filtered-probability signals are tradeable, and theses of this vintage frequently blur the line. The reported numbers also appear to ignore transaction costs, slippage, financing, dividends, and the practical impossibility of trading index levels directly, all of which erode multi-decade switching strategies substantially. Finally, three highly correlated US indices over overlapping periods is effectively one experiment, not three independent confirmations. The Kalman-filter regime-tracking idea is legitimate and the improved model fit may be real, but the trading-performance claim is almost certainly overstated and not credible as a forward-looking expectation without out-of-sample validation and cost-aware accounting.
Kalman Filter Demystified: from intuition to probabilistic graphical model to applications in finance#
This expository and methodological paper by Eric Benhamou (arXiv, 2018; 44 pages) revisits Kalman filter theory and rebuilds it from financial intuition up through a probabilistic graphical-model formulation, connecting the Kalman filter to Hidden Markov Models within a unified graphical framework. Beyond exposition, the paper makes methodological contributions: new inference algorithms for extended Kalman filters and the use of CMA-ES (Covariance Matrix Adaptation Evolution Strategy) optimization for parameter estimation in place of the traditional Expectation-Maximization (EM) approach. It then tests various dynamics assumptions for applying Kalman filters to market data.
Our summary: the paper’s value is twofold. As pedagogy, it offers an unusually careful path from intuition to graphical-model formalism, making explicit the Kalman-filter/HMM correspondence. As method, it proposes CMA-ES-based parameter estimation for (extended) Kalman filters, arguing this avoids some limitations of EM, and applies the resulting filter to build a trend-following technical-analysis system, reporting superior performance for trend-following detection versus conventional approaches.
Key metrics: this is primarily a tutorial/methodological paper, and it reports no rigorous quantitative trading metrics (no Sharpe ratio, hit rate, drawdown, or out-of-sample backtest statistics). The financial application is illustrated qualitatively as improved trend-following detection rather than through a formal performance table. No standardized error metrics (log-likelihood/AIC/BIC, RMSE), specific market/sample period, or code repository are stated in the accessible abstract. The concrete deliverables are the graphical-model exposition, the extended-Kalman inference algorithms, and the CMA-ES estimation procedure.
Critical view: as a reference and teaching document the paper is genuinely useful and the Kalman-filter/HMM/graphical-model unification is sound, well-trodden theory presented clearly. The methodological claim, that CMA-ES improves estimation over EM, is plausible (CMA-ES can escape local optima that trap EM) but the paper does not appear to back it with a controlled, statistically tested benchmark across many series. The “superior performance for trend-following” claim is the weakest part from an evaluation standpoint: trend-following backtests are notoriously easy to flatter through parameter choice, period selection, and ignored transaction costs, and without an out-of-sample protocol, multiple-asset validation, and cost accounting such a claim is descriptive rather than demonstrated. Readers should treat the trading result as an illustrative proof-of-concept, not as evidence of a deployable edge, while valuing the paper mainly for its clear derivations and the practical CMA-ES estimation recipe.
Advance Detection of Bull and Bear Phases in Cryptocurrency Markets#
This paper by Rahul Arulkumaran, Suyash Kumar, Shikha Tomar, Manideep Gongalla and Harshitha (arXiv, 2024) attempts to anticipate bull and bear phases in the cryptocurrency market by forecasting Bitcoin’s price path and then deriving regime labels from moving-average crossovers. Bull and bear phases are defined by the relationship between Bitcoin’s 50-day and 200-day moving averages (the classic golden-cross / death-cross convention). The authors forecast future Bitcoin prices with predictive algorithms and then compute the predicted 50-day and 200-day moving averages to flag upcoming phase transitions in advance. They build and compare two models: a Multiple Linear Regression (MLR) system and an LSTM neural network.
Our summary: the core idea is to detect regime changes early by predicting the moving averages that define them, rather than waiting for the crossover to occur. Bitcoin is used as the sole proxy for the whole market because of its roughly 50% dominance. Two architectures are compared: an MLR setup using 22 separate models to predict closing prices across a 21-day horizon, and an LSTM with 100 input neurons, hidden layers of 15 and 31 neurons, and 22 output neurons. The authors deliberately judge accuracy by visually comparing predicted versus actual SMA curves instead of using R-squared.
Key metrics: data are daily Bitcoin OHLCV series from an open API starting 1 January 2012; train/test split is 75/25 for both MLR and LSTM. The LSTM predicts 22 outputs over a 21-day horizon. Critically, the paper reports no conventional quantitative performance metrics: the authors explicitly decline to use R-squared (arguing the technical indicators are highly correlated) and report no RMSE, MAE, classification accuracy, precision/recall/F1, detection lead time, Sharpe ratio, or trading returns. Model quality is assessed qualitatively by overlaying predicted and actual SMA curves; they note the LSTM was less accurate at 1000 epochs than at 2000 epochs. No code repository is mentioned.
Critical view: this is a weak empirical paper by quantitative standards and its claims are essentially undemonstrated. Refusing to report any numerical accuracy metric, and substituting eyeball comparison of moving-average curves, makes the results impossible to verify or compare and is a major red flag. Predicting the 50-day and 200-day SMAs is also a deceptively easy task: these are heavily smoothed, highly autocorrelated quantities, so a model can appear to track them closely while carrying essentially no information about the underlying daily returns that actually matter for a crossover’s timing. Because the moving averages embed up to 200 days of past prices, “advance detection” risks being trivial near-persistence rather than genuine forecasting, and the paper provides no out-of-sample lead-time measurement, no comparison against a naive persistence baseline, and no trading backtest with transaction costs to show economic value. Using only Bitcoin and a single contiguous sample further limits generality. The conceptual framing (predict the indicators that define the regime) is reasonable, but as executed the work is descriptive and unvalidated, and its conclusions should be treated as suggestive at best.
Bayesian change point analysis of Bitcoin returns#
This Finance Research Letters paper by Sven Thies and Peter Molnar (2018) studies whether the average return and volatility of Bitcoin are stable over time. The authors apply a Bayesian change point (BCP) model to the daily Bitcoin return series to detect structural breaks and partition the time series into homogeneous segments. The BCP framework allows an unrestricted number of independent change points rather than imposing a prespecified count, so the data themselves dictate how many breaks occur in the first and second moments of the return distribution.
Our summary: the study finds that structural breaks in both the mean and volatility of Bitcoin returns are very frequent. Roughly forty-eight change points in the average return are detected, and segments with similar statistical properties are merged into about seven volatility regimes. Most regimes show positive average returns; one regime shows negative average returns. The novel takeaway is that, across regimes, higher volatility tends to accompany higher average return, with the notable exception of the most volatile regime, which is the only one delivering negative average returns, challenging a simple risk-return interpretation for Bitcoin.
Key metrics: approximately 48 change points in average return, merged into roughly 7 volatility regimes. The paper reports no trading or financial performance metrics (no Sharpe, strategy returns, drawdown, or hit rate); its outputs are purely descriptive distributional statistics (mean and volatility per regime). The analysis is on daily Bitcoin returns (Finance Research Letters vol. 27, pp. 223-227). No code or data availability is mentioned.
Critical view: this is an honest, modest descriptive study and does not overclaim. The BCP method is a reasonable, well-established Bayesian tool, and the finding of frequent breaks is plausible for an immature, highly volatile asset. The key caveat is that the entire analysis is in-sample and purely descriptive: change points are identified retrospectively over the full sample, so there is inherent look-ahead and no out-of-sample or real-time detection-lag assessment. The regimes therefore cannot be used directly for trading without a forward-looking detection scheme, and the paper makes no predictive or profitability claims, which is appropriate but limits practical applicability. The single-asset (Bitcoin-only) scope and short early-era sample also limit generality, and reproducibility is hampered by the absence of shared code. As descriptive evidence of non-stationarity it is solid; as a basis for strategy it is only suggestive.
Tensor time series change-point detection in cryptocurrency network data#
This 2025 arXiv preprint by Andreas Anastasiou and Ivor Cribben proposes TenSeg, a method for detecting multiple change points in tensor-valued (multi-network) time series, motivated by cryptocurrency fraud and market-manipulation detection across multiple trading platforms. Because manipulators increasingly operate across several interconnected venues simultaneously, the authors argue for analyzing a stack of trading networks jointly rather than each network in isolation. TenSeg works in two stages: first a tensor decomposition of the data, then detection of multiple change points in the cross-covariance structure of the decomposed components. The procedure is designed to handle frequent changes of possibly small magnitude and to be computationally fast.
Our summary: the contribution is methodological, extending change-point detection from vector/matrix time series to tensor time series by combining tensor decomposition with cross-covariance segmentation. The method is validated on simulated data and applied to real Ethereum blockchain tensor-variate trading network data. The authors report that TenSeg substantially outperforms existing state-of-the-art change-point techniques, and they release code on GitHub.
Key metrics: the abstract reports no concrete numerical performance figures (no detection accuracy, false-positive/false-negative rates, F1, or runtime numbers), and no trading or financial performance metrics (returns, Sharpe, drawdown); this is a statistical change-point methods paper, not a trading study. Evaluation is via simulation studies plus an Ethereum network application. Code is stated to be available on GitHub, aiding reproducibility; the specific sample period and number of networks/assets are not given in the available abstract.
Critical view: as a methods contribution this looks credible: tensor change-point detection is a genuine and underdeveloped problem, the two-stage decompose-then-detect design is principled, and simulation benchmarking against competing methods plus released code are good signs for reproducibility. The main caveats are that the “substantially outperforms” claim cannot be assessed without the actual numbers and the choice of competitor baselines, and simulation-based superiority can be sensitive to the data-generating process chosen. The applied claim, that detected change points correspond to manipulation or fraud, is largely descriptive/illustrative: detecting a structural break in a blockchain network is not the same as validating it against ground-truth fraud labels, which the abstract does not mention. There are no out-of-sample predictive or economic-value claims, so the practical value for trading or surveillance remains unproven. Overall a plausible methods paper whose real-data validation appears qualitative rather than label-validated.
Herding behavior in exploring the predictability of price clustering in cryptocurrency market#
This Finance Research Letters (2023, vol. 57) paper by Hachicha, Masmoudi, Abid and Obeid investigates price clustering, the tendency of traded prices to concentrate at particular digits, in the cryptocurrency market and links it to herding behavior. The authors focus on AI and big-data token markets. They use the Chang et al. (2000) cross-sectional absolute deviation (CSAD) model to test for static and time-varying herding, and they propose using K-means clustering together with a Hidden Markov Model (HMM) to characterize and predict the price-clustering phenomenon.
Our summary: the paper combines a classical behavioral-finance herding test (CSAD) with machine-learning/state-space tools (K-means and HMM) to study whether price clustering in crypto markets is predictable and whether it co-moves with herding. The novelty is the methodological pairing, applying clustering and HMM regime modeling to the price-clustering question in the relatively niche AI/big-data token segment, and the linkage of clustering patterns to herding dynamics that vary over time.
Key metrics: the paper reports its CSAD herding estimates and uses K-means and HMM, but specific numbers (number of HMM states/clusters, sample period, number of tokens, predictive accuracy, AIC/BIC) are not surfaced in the accessible summary. No trading or financial performance metrics (returns, Sharpe, drawdown) are reported; the contribution is statistical/behavioral rather than a trading strategy. No data or code availability is mentioned.
Critical view: the topic is legitimate and the CSAD herding methodology is standard, but several concerns temper the results. Price clustering and herding studies are largely descriptive and prone to data-snooping when many digit-clustering definitions and token subsamples are tried; the choice of the narrow “AI and big-data token” universe raises sample-selection and survivorship concerns and limits generality to the broader crypto market. HMM/K-means “predictability” claims are easy to overstate because regime labels are often fit in-sample and may suffer detection lag and look-ahead bias if the full sample is used for both clustering and evaluation, and the abstract does not indicate genuine out-of-sample testing. Without reported predictive accuracy numbers, transaction-cost-aware backtests, or any economic-value assessment, the “predictability” finding should be read as in-sample association rather than demonstrated forecastability. Reproducibility is weak given no stated data window or code.
Cluster analysis on the structure of the cryptocurrency market via Bitcoin-Ethereum filtering#
This Physica A (2019, vol. 527, art. 121339) paper by Jung Yoon Song, Woojin Chang and Jae Wook Song analyzes the correlation structure of the cryptocurrency market using correlation-based agglomerative hierarchical clustering and minimum spanning trees (MST). The methodological novelty is a “Bitcoin-Ethereum filtering” step that removes the linear influence of Bitcoin and Ethereum on the other cryptocurrencies before clustering, isolating residual co-movement structure that is otherwise masked by the two dominant coins. The authors examine market structure across three windows (Total, Pre-regulation, and Post-regulation) to see how regulatory announcements from various countries reshaped the network.
Our summary: after filtering out Bitcoin and Ethereum, the analysis confirms the leadership/centrality of Bitcoin and Ethereum in the raw market and uncovers six homogeneous clusters among the relatively less-traded cryptocurrencies. The key finding is that the market’s structure transforms after regulatory announcements, with the cluster/MST topology changing between the pre- and post-regulation periods. The filtering approach is the main contribution, offering a cleaner view of secondary co-movement structure.
Key metrics: six homogeneous clusters are identified among less-traded cryptocurrencies; the analysis is split into Total, Pre-, and Post-regulation periods. The paper reports no trading or financial performance metrics (no returns, Sharpe, drawdown, or hit rate); outputs are network/structural statistics (correlation-based hierarchical clustering and MST topology). Exact number of cryptocurrencies and precise sample dates are not surfaced in the accessible abstract (DOI: 10.1016/j.physa.2019.121339). No code availability is mentioned.
Critical view: this is a sound, descriptive econophysics study and does not overclaim predictive power. Correlation-based MST and hierarchical clustering are well-established tools, and the Bitcoin-Ethereum linear-filtering idea is a sensible, interpretable refinement. The principal caveats are inherent to the genre: results are entirely in-sample and structural, with no out-of-sample or predictive validation, so the regulatory “transformation” is an association, not a causal or forecastable effect; the pre/post split is researcher-chosen and vulnerable to confounding with broader market moves (e.g., the 2017-2018 boom-bust) rather than regulation per se. Correlation networks are also sensitive to estimation window, the linear-filter assumption ignores nonlinear dependence, and the number of clusters can depend on linkage/threshold choices. No economic-value claims, transaction costs, or trading evaluation are present, and reproducibility is limited by the absence of a stated coin list and shared code. As a structural snapshot it is credible; as actionable insight it is only suggestive.
A Deep Learning Framework for Predicting Digital Asset Price Movement from Trade-by-trade Data#
This 2020 arXiv preprint by Qi Zhao presents an LSTM-based deep learning framework for forecasting short-term cryptocurrency price movements directly from granular trade-by-trade (tick-level) data rather than aggregated OHLCV bars. The approach emphasizes feature engineering over high-frequency market-microstructure information and extensive hyperparameter optimization, training LSTM networks to predict the direction of price moves within defined short time windows. The dataset spans roughly one year of trade-by-trade cryptocurrency data.
Our summary: the framework reports over 60% out-of-sample directional accuracy on unseen trading periods and, notably, claims transferability, that a model trained on some cryptocurrencies generalizes to other coins not in the training set, suggesting the LSTM captures somewhat universal microstructure patterns. The author further claims profitability in a “realistic trading simulation” setting. The novelty is the use of raw trade-by-trade data plus the cross-asset transfer result.
Key metrics: out-of-sample directional accuracy is reported as “over 60%”; the dataset is ~1 year of trade-by-trade data; the paper claims profitability in a trading simulation and cross-cryptocurrency transfer. No specific Sharpe ratio, net returns, drawdown, turnover, or transaction-cost figures are given, and the exact accuracy, number of assets, and sample dates are not precisely stated. No data or code availability is mentioned.
Critical view: the headline numbers warrant skepticism. A “60%+ directional accuracy” on high-frequency crypto data is impressive on its face but is highly sensitive to label construction (very short horizons have strong autocorrelation and can be near-trivially predictable in-sample), class imbalance, and whether the baseline is properly chosen, accuracy alone is a weak metric without a no-skill benchmark. The crucial omission is transaction costs and the bid-ask spread: at tick frequency, spreads and fees typically erase directional edges far smaller than the implied signal, so the “realistic trading simulation” profitability claim cannot be assessed without explicit cost assumptions, slippage, and execution modeling, none of which are documented. Single-author preprint status (not peer-reviewed), unspecified sample/asset details, and no released code make reproducibility and data-snooping/overfitting hard to rule out, extensive hyperparameter search on one year of one market is a classic overfitting risk. The cross-asset transfer result is the most interesting and least cost-dependent claim, but it too needs out-of-sample, out-of-period verification. Treat the results as preliminary and likely optimistic until independently reproduced with costs.
Designing a cryptocurrency trading system with deep reinforcement learning utilizing LSTM neural networks and XGBoost feature selection#
This Applied Soft Computing (2025, vol. 175, art. 113029) paper by H. Ghadiri and E. Hajizadeh designs a two-stage cryptocurrency trading system. First, XGBoost is used for feature selection, ranking the most relevant inputs from a large pool spanning market variables, technical indicators, macroeconomic factors, and blockchain-specific (on-chain) data, separately for each cryptocurrency. Second, the selected features are fed into a Double Deep Q-Network (DDQN) reinforcement-learning agent whose function approximator incorporates LSTM, BiLSTM, and GRU recurrent layers to emit discrete trading signals (buy, hold, sell). The system is evaluated on Bitcoin and Ethereum.
Our summary: the central claims are that (i) blockchain/on-chain variables carry crucial information for trading decisions and survive feature selection, and (ii) coupling XGBoost feature selection with the recurrent DDQN agent improves all key trading performance metrics relative to using the DDQN alone. The novelty is the explicit XGBoost-driven feature-selection front end combined with a recurrent (LSTM/BiLSTM/GRU) deep-RL agent and the inclusion of on-chain data as inputs.
Key metrics: tested on Bitcoin and Ethereum over July 2021 to March 2023. The paper reports that XGBoost feature selection improves “all key trading performance metrics” of the DDQN agent, but specific numerical values (cumulative/annualized return, Sharpe ratio, maximum drawdown, win rate, transaction costs) are not surfaced in the accessible abstract. No explicit statement on data or code availability appears in the abstract.
Critical view: several red flags common to deep-RL trading papers apply. The evaluation window (July 2021-March 2023) is short (under two years) and covers a specific, largely bearish/sideways crypto regime, so reported performance may not generalize; with only two assets the effective out-of-sample breadth is very small and overfitting/data-snooping risk is high, especially given a large initial feature pool fed through XGBoost selection (selection itself can leak look-ahead information if performed on the full sample rather than walk-forward). Deep-RL agents are notoriously sensitive to seed, hyperparameters, and reward shaping, and “improves all key metrics” is a relative claim against the authors’ own baseline rather than against a buy-and-hold or simple momentum benchmark, which is the more honest comparison. Whether transaction costs, slippage, and realistic execution are included is critical for any discrete buy/hold/sell agent and is not confirmed in the abstract. Without reported absolute Sharpe/drawdown numbers, a proper walk-forward protocol, multiple-asset breadth, and released code, the results should be considered promising but unverified and at meaningful risk of being overstated.
Feature selection in jump models#
Peter Nystrup, Petter N. Kolm, and Erik Lindström (Lund University, NYU Courant, and Technical University of Denmark; Expert Systems with Applications 184, 2021, article 115558) extend the jump model with joint feature selection, producing the sparse jump model (SJM) that the crypto and portfolio applications later build on. Feature selection is essential in high-dimensional settings where the number of candidate features is large relative to the number of observations and the states differ only with respect to a subset of features. The authors develop a coordinate-descent algorithm that alternates between weighting features and estimating the model parameters and state sequence, scaling to large data sets with many noisy features.
Our summary: where the original jump model assumed the analyst had already chosen the right features, this paper makes feature selection part of the estimation, so the model itself reveals which features separate the regimes. By leveraging information embedded in the ordering of the data, the sparse jump model recovers true features and rejects irrelevant ones with high probability even among hundreds of candidates, and it does so without supervised labels. The result is more accurate and more persistent state estimates than competing methods, and a model that is remarkably robust to noise — the property that makes large-scale feature attribution (as in the cryptocurrency paper below) feasible.
Data and code: two simulated and three real data sets — financial returns (from Kenneth French’s data library), protein sequences, and Wikipedia text — chosen to stress different separation structures. A Python implementation is provided online as supplementary material; benchmarks again use hmmlearn.
Key metrics: on a high-dimensional text-segmentation task with P = 300 true features, the sparse jump model located breakpoints with a mean absolute offset of only 0.48 words, versus 3.5 for the Greedy Gaussian Segmentation (GGS) benchmark and 35 for the standard jump model — a striking demonstration that feature weighting, not just jump penalization, drives accuracy in high dimensions. The financial-returns application showed the framework also works when the information separating states lives in the second moment (volatility regimes).
Critical view: this is a methods paper, and its empirical claims are about recovery accuracy and robustness rather than economic performance, so there is no Sharpe/drawdown evidence here. The same forward-looking-feature caveat from the 2020 paper applies, and the authors explicitly leave principled hyperparameter selection (the number of effective features and the jump penalty) to cross-validation or backtesting in future work, which means downstream users still carry tuning risk. As the statistical backbone of the sparse jump model it is well validated; as a trading result it makes no claims.
What drives cryptocurrency returns? A sparse statistical jump model approach#
Federico P. Cortese, Petter N. Kolm, and Erik Lindström (University of Milano-Bicocca, NYU Courant, and Lund University; Digital Finance 5, 2023, pp. 483–518) apply the sparse statistical jump model to identify the features that drive the return dynamics of the largest cryptocurrencies. The algorithm jointly performs feature selection, parameter estimation, and state classification over a large set of candidate features spanning cryptocurrency-, sentiment-, and financial-market-based time series, some drawn from the emerging crypto literature and some newly proposed by the authors. This is the most directly relevant precedent for applying jump models to crypto regime detection.
Our summary: feeding more than four hundred candidate features into a sparse jump model, the authors find that a three-state specification best describes crypto returns, with the states carrying natural bull, neutral, and bear interpretations. The data-driven feature selection then answers the title’s question: the key drivers are the first moments of returns (exponential moving averages), trend and reversal signals from the technical-analysis literature, and measures of market activity and public attention — the last two being the paper’s novel feature contributions. Notably, second moments (volatility) are not among the selected drivers, distinguishing this result from volatility-regime studies. The work demonstrates that the SJM’s robustness to noise (established in Feature selection in jump models) makes large-scale, interpretable feature attribution practical in a notoriously noisy market.
Data and code: daily data for the five largest and most liquid cryptocurrencies — Bitcoin (BTC), Ethereum (ETH), Ripple (XRP), Litecoin (LTC), and Bitcoin Cash (BCH) — over January 2018 to September 2022, with more than 400 candidate features. The article is open access. It uses the sparse jump model framework subsequently packaged as jumpmodels.
Key metrics: the empirical result is a three-state regime model selected over a 400+ feature set, with a sparse subset of features (return EMAs, trend/reversal technical signals, market-activity and public-attention proxies) retained as the relevant drivers. Because the paper is a feature-attribution and regime-identification study rather than a strategy backtest, it does not report trading metrics such as annualized return, Sharpe ratio, or maximum drawdown; its evaluation is the interpretability and economic plausibility of the selected features and inferred regimes.
Critical view: the contribution is descriptive — which features matter and how many regimes — rather than a tradable alpha claim, and the authors frame it as such, noting only that practitioners could use the identified features to distinguish trends and detect regime switches. The usual cautions apply: feature selection performed over the full sample can leak look-ahead information, regime labels are assigned post hoc, and the single 2018–2022 window covers one particular boom-bust cycle, so the selected feature set may not be stable out of sample. As an interpretable mapping of crypto regime drivers it is credible and useful; it should not be read as evidence that a jump-model crypto strategy is profitable.
Identifying patterns in financial markets: extending the statistical jump model for regime identification#
Afşar Onat Aydınhan, Petter N. Kolm, John M. Mulvey, and Yizhan Shu (Princeton University and NYU Courant; Annals of Operations Research, 2024) extend the statistical jump model into a continuous jump model by generalizing the discrete hidden state variable into a probability vector over all regimes. Instead of committing to a single active state at each time, the model estimates the probability of being in each regime, which provides richer information for downstream tasks such as regime-aware portfolio construction and risk management. The smooth transition from one regime to another is designed to improve robustness over the original discrete model.
Our summary: this is the “Continuous JM” that softens the hard state assignment of the original (Learning hidden Markov models with persistent states by penalizing jumps) and sparse (Feature selection in jump models) jump models into probabilistic regime membership. The authors give a probabilistic interpretation of the continuous model and introduce a novel penalty term — mode loss — that pushes the probability estimates toward the vertices of the probability simplex, sharpening regime identification while retaining the smoothness benefits. The headline claim is that the continuous formulation outperforms traditional regime-switching models, with the advantage most pronounced precisely where classical methods struggle: when regimes are imbalanced and historical data is limited, both endemic to financial markets.
Data and code: an extensive simulation study (two- and three-state HMM data-generating processes, including misspecified t-distributed emissions and non-geometric sojourn times) plus an application to daily log-returns of the Nasdaq Composite index across two contrasting 10-year periods — 1996–2005 (the dot-com bubble and burst) and 2013–2022 (a relatively calm regime). The models use only backward-looking features (rolling mean returns and volatilities over windows of 6 and 14 days), deliberately chosen so the estimator supports online, out-of-sample prediction. The continuous model discretizes the probability simplex on a uniform grid (size 0.05). The reference implementation is the jumpmodels Python package maintained by co-author Yizhan Shu. (Working-paper version, March 2024; the published version appears in Annals of Operations Research.)
Key metrics: in the three-state simulation the continuous jump model improves on the discrete jump model by roughly 2% in balanced accuracy (BAC) and 4% in ROC-AUC; more generally it attains BAC comparable to the discrete model while delivering markedly better ROC-AUC, i.e. more reliable per-regime probability estimates — the property that makes the soft assignments useful for downstream regime-aware portfolio and risk tasks. The mode-loss penalty gives a further small improvement in BAC at short sample sizes by promoting sparsity/persistence in the estimated probability vectors. Both jump models clearly outperform Baum-Welch-estimated HMMs, with the gap widest when regimes are imbalanced, persistent, and the history is short. The main cost is compute: enumerating the simplex grid raises running time by a factor on the order of hundreds relative to the discrete model.
Critical view: having now read the full paper (retrieved via the open SSRN working-paper version), the contribution holds up — probabilistic regime membership plus a simplex-vertex mode-loss penalty, with a clean optimal-transport link back to the discrete jump penalty, is a principled generalization, and the simulation design honestly stresses misspecification and small-sample/imbalanced regimes where the gains are claimed. Two caveats temper the practical reading. First, the headline advantage is in probability calibration (ROC-AUC) rather than hard-label accuracy (BAC is only ~2% better than the already-strong discrete model), so the payoff depends on actually using the probabilities downstream — the paper demonstrates better estimates but stops short of a portfolio or risk backtest, which it leaves to future work. Second, the hundreds-fold compute overhead is a real deployment consideration for large state spaces. As a regime-identification method it is well validated; its economic value awaits the portfolio application the authors flag.
Bayesian Online Changepoint Detection#
This paper by Ryan Prescott Adams and David J.C. MacKay (2007, arXiv:0710.3742) is the foundational reference for online, Bayesian regime-change detection and one of the most widely implemented algorithms in the streaming-data toolkit. Assuming the data-generating parameters before and after a changepoint are independent, the authors derive an exact online algorithm that maintains the posterior distribution over the current “run length” — the time elapsed since the most recent changepoint — updated with each new observation via a simple message-passing recursion. Because it is causal and recursive, it is directly suited to live market feeds: at every tick you get a full posterior over “how long has the current regime lasted,” and a spike in changepoint probability flags a regime break in real time.
Our summary: this is the Bayesian counterpart to the statistical jump models elsewhere in this collection — both answer “has the regime just changed?” but BOCD does it with an explicit, exact posterior over run length rather than a penalized clustering objective. For a systematic trader it is a small, beautiful, directly implementable algorithm (well-supported in open-source libraries) that you can wire to volatility, spread, or correlation series to gate strategy parameters. The one practical caveat is computational: vanilla BOCD’s cost grows with the number of run-length hypotheses, so streaming deployment needs pruning/truncation of the run-length distribution. Essential full read; pair it with a heavy-tailed observation model for crypto.
Data and code: a methods paper with illustrative examples (including financial data); finance is explicitly named as an application. Reference implementations exist in R (“ocp”) and many open-source Python libraries.
Key metrics: this is an inference-method paper, not a trading backtest — there are no Sharpe/return figures. Its value is the exact run-length posterior and the O(run-length) message-passing recursion that makes online regime detection tractable.
Particle Filters for Markov-Switching Stochastic Volatility Models#
This paper by Yong Bao, Carl Chiarella and Boda Kang (2012, SSRN 2163902; later a chapter in the Oxford Handbook of Computational Economics and Finance, OUP 2018) tackles sequential Bayesian estimation of regime-switching stochastic-volatility models — covering both regime detection and the particle-filter/sequential-Monte-Carlo machinery in one place. The authors build an auxiliary particle filter (ASIR) for Markov-switching stochastic volatility that estimates the regime transition probabilities online via a continually updated Dirichlet distribution, so both the latent volatility state and the switching dynamics are learned sequentially as data arrive.
Our summary: the key practical result is about when particle methods stay accurate. Their stated improvement over the earlier Carvalho-Lopes (2006/2007) approach — Liu-West kernel smoothing combined with the Pitt-Shephard auxiliary particle filter — is accuracy when the probability of switching from one regime to a different regime is high. Kernel smoothing degrades in exactly that high-transition setting, whereas the Dirichlet/APF combination accommodates it. The authors note such frequent regime flips are common in energy, commodity and FX markets — and the same is conspicuously true of crypto, which makes this a relevant template for regime-aware volatility estimation there. Methodological/theoretical; read it for the filter construction (and the honest discussion of effective sample size in Table 4) rather than a trading track record.
Data and code: methodological study with simulation and empirical illustrations; an open working-paper version is hosted by UTS (QFR research paper 299). No public code.
Key metrics: the paper reports filtering accuracy and effective-sample-size comparisons against the kernel-smoothing benchmark rather than trading P&L; the headline is robustness under high regime-transition probabilities.