Model-Based Feature Extraction

David A. Dickey and Wayne A. Fuller (1979) — Journal of the American Statistical Association · 27997 citations

Mark B. Garman and Michael J. Klass (1980) — The Journal of Business · 1472 citations

Introduces a set of volatility estimators using Open, High, Low, and Close prices that are up to 8 times more efficient than standard close-to-close variance calculations.

Robert F. Engle (1983) — Journal of Money, Credit and Banking · 503 citations

This seminal paper introduces the Autoregressive Conditional Heteroscedasticity (ARCH) model to estimate the time-varying conditional variance of U.S. inflation, revealing that high inflation does not necessarily imply high unpredictability.

Tim Bollerslev (1986) — Journal of Econometrics · 23212 citations

This paper introduces the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, a significant extension of the ARCH model that allows for more flexible and parsimonious modeling of time-varying volatility by incorporating past conditional variances.

Robert F. Engle and C. W. J. Granger (1987) — Econometrica · 31736 citations

Engle and Granger (1987) formalize cointegration and prove that cointegrated I(1) variables must admit an error-correction representation, then provide practical two-step estimation and simulation-based cointegration tests with empirical macro/finance examples.

James D. Hamilton (1989) — Econometrica · 9717 citations

Hamilton (1989) introduces a maximum-likelihood Markov-switching autoregressive framework and nonlinear filter to infer unobserved regime changes, and shows U.S. real GNP growth is well-described by recurrent expansion/recession regimes with recessions implying an ~3% permanent level loss.

Søren Johansen and Katarina Juselius (1990) — Oxford Bulletin of Economics and Statistics · 11542 citations

This paper presents a maximum likelihood approach for estimating and testing cointegration relationships in vector autoregressive (VAR) models, with a focus on linear restrictions on cointegration vectors and weights, and illustrates the method using money demand data from Denmark and Finland.

Daniel B. Nelson (1991) — Econometrica · 10571 citations

This paper introduces Exponential GARCH (EGARCH) to address limitations of standard GARCH models, such as the inability to capture the negative correlation between returns and volatility, restrictive parameter constraints, and difficulties in interpreting volatility persistence.

Søren Johansen (1991) — Econometrica

This paper derives maximum-likelihood estimation and likelihood-ratio tests for cointegration rank and cointegration-vector restrictions in Gaussian VARs, yielding the “Johansen tests” based on a canonical-correlation eigenvalue problem with nonstandard asymptotic critical values.

Lawrence R. Glosten et al. (1993) — Journal of Finance · 9180 citations

Dennis Yang and Qiang Zhang (2000) — The Journal of Business · 470 citations

The authors introduce the 'Yang-Zhang' volatility estimator, which uses OHLC data to provide a minimum-variance estimate that is robust to both price trends (drift) and overnight gaps (opening jumps).

R. Cont (2001) — Quantitative Finance · 3429 citations

This paper reviews the 'stylized facts' of asset returns across various markets and instruments, highlighting statistical properties like heavy tails, volatility clustering, and the absence of linear autocorrelations, while also discussing the statistical issues that arise when analyzing financial time series.

International Asset Allocation With Regime Shifts

Andrew Ang and Geert Bekaert (2002) — Review of Financial Studies · 1567 citations

Despite correlations rising in bear markets, international diversification remains economically valuable, particularly when investors can switch into cash (risk-free assets) during high-volatility regimes.

Robert J. Elliott et al. (2005) — Quantitative Finance

Provides a concrete state-space (Kalman filter + EM estimation) framework for pairs trading by modeling the spread as a mean-reverting Gaussian process observed with noise and using deviations from the filtered/predicted spread to trigger trades.

Fulvio Corsi (2009) — Journal of Financial Econometrics · 2370 citations

This paper introduces the Heterogeneous Autoregressive model of Realized Volatility (HAR-RV), a simple additive cascade model using volatility components across different time horizons, which effectively replicates key empirical features of financial returns like long memory and fat tails, while also demonstrating strong forecasting performance.

Andrew Ang and Allan Timmermann (2011) · 454 citations

This paper reviews how regime-switching models (HMMs) capture the abrupt, persistent changes in financial data (volatility clustering, skewness) that linear models miss, demonstrating that optimal portfolios must dynamically adjust to 'bull' and 'bear' states.

Matthew D. Hoffman and Andrew Gelman (2011) — arXiv:1111.4246 [cs, stat] · 4962 citations

This paper introduces NUTS, an extension of Hamiltonian Monte Carlo that automatically chooses trajectory length (and adaptively tunes step size), delivering HMC-level efficiency without hand-tuning.

Jim Gatheral et al. (2014) · 827 citations

This paper demonstrates that log-volatility behaves like a fractional Brownian motion with a Hurst exponent around 0.1, leading to the Rough FSV model, which aligns well with financial data and improves volatility forecasting.

Alan Moreira and Tyler Muir (2017) — The Journal of Finance · 381 citations

Scaling factor exposures each month by the inverse of last month’s realized variance produces large alphas and materially higher Sharpe ratios across many factors because volatility forecasts risk much more than it forecasts expected returns.

Advances in Financial Machine Learning

Marcos Lopez de Prado (2018) — John Wiley & Sons · 106 citations

Michael Betancourt (2018) — arXiv:1701.02434 [stat] · 1393 citations

This tutorial explains Hamiltonian Monte Carlo (HMC) through the geometry of the “typical set,” showing why gradient-informed, energy-conserving trajectories can explore high-dimensional posteriors far more efficiently—and how tuning/diagnostics (mass matrix, step size, trajectory length, divergences) make or break performance.

A. Sinem Uysal and John M. Mulvey (2021) — The Journal of Financial Data Science · 20 citations

The paper uses supervised ML (especially random forests) to predict recessions and equity “crash” regimes from macro data and then uses these probabilities to improve risk parity portfolios via regime-aware covariance estimation and overlay trades.

Blanka Horvath et al. (2021) · 12 citations

This paper introduces the Wasserstein k-means (WK-means) algorithm, a robust, non-parametric method for clustering financial time series into distinct market regimes by treating segments as probability distributions and using the p-Wasserstein distance, outperforming traditional moment-based k-means and HMMs, especially for non-Gaussian data.

Ailing Zeng et al. (2022) · 3371 citations

This paper challenges the widespread adoption of Transformer models for long-term time series forecasting (LTSF), demonstrating that simple linear models consistently outperform them across various real-world datasets.

Kevin Khang (2022) — The Journal of Portfolio Management · 1 citations

Risk-model volatility forecasts depend heavily on the estimation lookback window, and a practical way to become “regime-aware” is to monitor real-time dispersion across forecasts from different calibrations and switch between slow- and fast-moving models when misspecification risk is high.

Stephen Marra (2023) — The Journal of Portfolio Management

A practitioner-focused survey comparing common volatility forecasting models (historical, ARMA/GARCH, and option-implied) and showing why relatively simple, well-designed historical models can be robust inputs for volatility targeting and risk-parity allocation.

Sophia Sun and Rose Yu (2025) · 1 citations

This paper introduces Conformal Prediction for Time-series with Change points (CPTC), a novel algorithm that integrates a model to predict underlying states with online conformal prediction to provide uncertainty quantification for time series data with change points, demonstrating improved validity and adaptivity compared to state-of-the-art baselines.

Unknown (2025)

Yizhan Shu and John M. Mulvey (2025) — The Journal of Portfolio Management · 2 citations

A dynamic factor allocation strategy using Sparse Jump Models (SJM) to identify active return regimes improves the Information Ratio from 0.05 to ~0.45 compared to an equal-weighted benchmark.

Ilya Chevyrev et al. (2026) — Springer Nature Switzerland · 260 citations

This paper introduces the signature method, a way to transform time-ordered data into a set of features using iterated integrals, and discusses its theoretical properties and machine learning applications, including handwritten digit classification.

J. Scott Armstrong and Fred Collopy — International Journal of Forecasting

The paper empirically compares common forecast error metrics and shows that RMSE is too unreliable for comparing methods across many time series, recommending relative/robust alternatives like GMRAE, MdRAE, and MdAPE depending on the task and sample size.

Michael Parkinson — The Journal of Business · 1938 citations

This paper introduces the extreme value method for estimating the variance of the rate of return of a common stock, demonstrating its superior efficiency compared to the traditional method using closing prices.

Rob Hyndman · 4190 citations

C. W. J. Granger (1980) — Journal of Economic Dynamics and Control

Granger proposes an operational, forecasting-based definition of causality for time series and argues that causal claims should be validated using out-of-sample forecasting improvements rather than in-sample model fit.

Jun Shao (1993) — Journal of the American Statistical Association

Shao shows that the ubiquitous leave-one-out cross-validation (and AIC/Cp/related criteria) is asymptotically inconsistent for linear-model variable selection, and that consistency requires an “opposite” split where the validation set is most of the data (n_v/n → 1) while the training set still grows (n_c → ∞).

M. Dai et al. (2010) — SIAM Journal on Financial Mathematics

This paper mathematically proves that a trend-following strategy based on two probability thresholds (buy/sell) is optimal in a regime-switching market and outperforms buy-and-hold by ~2x in historical tests.

Andrew J. Patton (2011) — Journal of Econometrics · 1172 citations

This paper examines how the use of imperfect volatility proxies, such as squared returns or realized volatility, can distort the comparison of conditional variance forecasts and derives necessary and sufficient conditions for loss functions to be robust to noise in the volatility proxy.

Nicolas Huck and Komivi Afawubo (2015) — Applied Economics

Using S&P 500 constituents from August 2000 to September 2011, the paper shows that cointegration-based pair selection delivers high, stable, and transaction-cost-robust pairs-trading alpha, while the popular distance method does not.

Gautier Marti et al. (2016)

This paper establishes the statistical consistency of clustering financial time series based on correlations and empirically determines that a window of 250 daily returns (approximately one year) with the Ward + Spearman methodology is sufficient for reliable cluster recovery for 265 assets, addressing the trade-off between spurious clusters from short data and smoothed dynamics from long data.

Ari Levine and Lasse Heje Pedersen (2016) — Financial Analysts Journal

This paper proves that Time-Series Momentum (TSMOM) and Moving-Average Crossover (MACROSS) strategies are mathematically equivalent linear filters, suggesting practitioners should focus on portfolio construction rather than signal selection.

John M. Mulvey and Han Liu (2016) — The Journal of Portfolio Management

The paper shows how a simple, data-driven two-regime (growth vs. contraction) model—detected via trend filtering on the S&P 500—materially increases estimated left-tail risk for endowments, and how contingent (drawdown-triggered) spending cuts can significantly reduce long-horizon downside risk.

J.D. Hamilton (2016) — Elsevier

A comprehensive survey of Markov-switching models for identifying economic regimes (recessions/expansions), detailing estimation via EM algorithms and Gibbs sampling.

Christopher Krauss (2017) — Journal of Economic Surveys

A structured survey of pairs trading/statistical arbitrage research that compares five major modeling frameworks and highlights what actually survives transaction costs and other market frictions.

Matthew Clegg and Christopher Krauss (2018) — Quantitative Finance

The paper upgrades classical cointegration pairs trading by modeling spreads as a mixture of a mean-reverting component and a random walk (“partial cointegration”), and shows this produces ~12% annualized net returns on S&P 500 constituents (1990–2015) while standard distance/cointegration variants largely fail on the same dataset.

Regina Hammerschmid and Harald Lohre (2018) — International Review of Economics & Finance

By applying Markov regime switching models to a PCA-derived macroeconomic factor, this paper demonstrates that regime-based signals provide complementary predictive power to fundamental and technical indicators, yielding a 3.77% annual utility gain for S&P 500 investors.

S. Endres and J. Stübinger (2019) — Applied Economics

The paper derives closed-form optimal entry/exit thresholds for pairs trading when spreads follow a Lévy-driven Ornstein–Uhlenbeck process, and reports very large net returns on 1-minute S&P 500 data (1998–2015), especially in sectors with strong mean reversion, volatility, and jump activity.

Peter Akioyamen et al. (2020) — Proceedings of the First ACM International Conference on AI in Finance

The paper proposes a PCA + k-means + classifier pipeline to detect U.S. macro “regimes” (crisis vs non-crisis) and shows how regime predictions can drive profitable tail-hedging and tactical allocation strategies out-of-sample (2014–2020).

Bryan Lim and Stefan Zohren (2021) — Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences

A compact survey of modern deep-learning architectures for time-series forecasting—covering CNN/RNN/attention designs, multi-horizon forecasting setups, hybrid statistical–neural models, and extensions for interpretability and counterfactual decision support.

Shereen Elsayed et al. (2021)

A carefully windowed and feature-engineered gradient-boosted tree baseline (XGBoost) matches or beats many recent deep-learning time-series forecasters across multiple public datasets, suggesting complexity is often unnecessary.

Gautier Marti et al. (2021) — Springer International Publishing · 137 citations

A comprehensive review of correlation-based clustering and network methods (MST, PMFG, hierarchical clustering, RMT, etc.) for understanding market structure, building portfolios, and monitoring systemic risk—along with why these methods are often unstable and hard to operationalize.

Anastasios N. Angelopoulos and Stephen Bates (2022) · 944 citations

This tutorial explains how conformal prediction turns any black-box model’s heuristic uncertainty into prediction sets/intervals with finite-sample, distribution-free coverage guarantees (e.g., 90% coverage using a small calibration set).

Sebastien Lleo et al. (2022) — The Journal of Portfolio Management

The paper applies Bayesian changepoint (mean/drift-switch) detection to time exits and re-entries (including long–short flips) in “bubblelike” equity price series, and shows large improvements over buy-and-hold in episodes like AAPL and the S&P 500 during early 2020.

Ali Shojaie and Emily B. Fox (2022) — Annual Review of Statistics and Its Application · 458 citations

This review explains what Granger causality really measures (predictive, time-ordered dependence), why naive/bivariate VAR tests can mislead, and how modern high-dimensional, nonlinear, discrete-valued, and mixed-frequency methods extend the framework.

Megan Czasonis et al. (2023) — The Journal of Portfolio Management

The paper introduces “relevance-based prediction,” a forecasting framework that weights historical observations by how informative and similar they are to today, and provides a prediction-specific reliability measure (fit) plus an adaptive procedure for jointly choosing data subsets and variables.

Azul Garza and Max Mergenthaler-Canseco (2023)

TimeGPT proposes a large pre-trained Transformer “foundation model” for time-series forecasting that can be used zero-shot on new datasets, aiming to match or beat strong baselines while being dramatically faster and simpler to deploy.

Qingsong Wen et al. (2023)

A comprehensive survey of how Transformer architectures are adapted for time series tasks (forecasting, anomaly detection, classification), plus empirical diagnostics showing where time-series Transformers struggle (long inputs, deeper models) and what consistently helps (seasonal-trend decomposition).

Irene Aldridge (2023) — The Journal of Portfolio Management

The article argues that “true AI” for finance is best thought of as unsupervised learning (e.g., SVD/PCA), which can extract dominant return drivers while avoiding key pitfalls of linear regression (collinearity, hypothesis bias) and the opacity of transformer-style models like ChatGPT.

Abdul Fatir Ansari et al. (2024)

Chronos introduces a minimalist framework that tokenizes time series data and trains off-the-shelf transformer-based language models for state-of-the-art probabilistic forecasting, demonstrating strong zero-shot capabilities and simplifying forecasting pipelines.

Tobias J. Moskowitz et al. (2025)

This paper demonstrates that a theoretically-motivated nonlinear transformation of past returns significantly improves time series momentum strategies across various asset classes and frequencies, particularly by enhancing tail-hedging properties during market downturns.

Nouha Karaouli et al. (2025)

Time Series Foundation Models (TSFMs) exhibit zero-shot capabilities highly dependent on their pretraining domains and often do not outperform smaller, specialized models when fine-tuned on unseen, domain-shifted real-world data, despite their larger parameter counts.

Xueying Ding et al. (2025)

Delphyne is a transformer time-series foundation model pre-trained on LOTSA plus finance data that argues cross-domain pre-training causes negative transfer in zero-shot, so the real value is fast few-step fine-tuning—especially for financial forecasting and risk tasks.

Yi Zuo and Eisuke Kita — Expert Systems with Applications

The paper forecasts stock valuation (P/E ratio) by discretizing it into clusters and learning a Bayesian network over lagged states, reporting higher accuracy than AR/MA/ARMA/ARCH—especially when discretization uses Ward clustering rather than uniform binning.

Rob J. Hyndman and Anne B. Koehler — International Journal of Forecasting

The paper shows that many popular forecast-accuracy metrics (especially MAPE/sMAPE and several M-competition/M3 measures) break down with zeros, negatives, or flat stretches, and proposes MASE as a robust, interpretable default for comparing forecast accuracy across many time series.

Jesús Gonzalo and Tae-Hwy Lee — Journal of Econometrics

The paper shows that Johansen’s likelihood-ratio cointegration tests can falsely detect cointegration with probability approaching 1 under common trend/memory and covariance misspecifications, while Engle–Granger is often more robust—so practitioners should not rely on unit-root pretests alone and should cross-check with both approaches.

Wen Long et al. — Knowledge-Based Systems

The paper proposes an end-to-end “multi-filters neural network” (MFNN) that combines CNN-style temporal convolutions and RNN-style recurrence to learn features from 1-minute OHLCVA sequences and predict extreme up/down moves of the CSI 300 index, improving both classification accuracy and trading-simulation profitability versus single-architecture networks and classic ML/statistical baselines.

A. Hyvärinen and E. Oja — Neural Networks

This invited review explains Independent Component Analysis (ICA)—how to recover statistically independent latent sources from linear mixtures—derives key objective functions (non-Gaussianity, mutual information, maximum likelihood), and presents the FastICA algorithm with example applications including finance.

G. William Schwert (1989) — The Journal of Finance · 3828 citations

A seminal analysis of 130 years of data showing that stock volatility rises significantly during recessions and with trading volume, but the extreme volatility of the Great Depression remains an anomaly unexplained by standard macroeconomic or leverage models.

Backpropagation through time: what it does and how to do it

P.J. Werbos (1990) — Proceedings of the IEEE · 5220 citations

The paper explains how to compute exact gradients efficiently for dynamic/recurrent systems using backpropagation through time (BPTT), with equations and pseudocode that generalize basic backprop to time-lagged networks and control/identification problems.

George Casella and Edward I. George (1992) — The American Statistician

Casella and George (1992) explain—using simple, fully worked examples—why the Gibbs sampler produces draws from a target marginal distribution by iterating draws from easier conditional distributions, and when it can fail.

LSTM can solve hard long time lag problems

Sepp Hochreiter and Jürgen Schmidhuber (1996) — MIT Press · 1063 citations

This paper introduces Long Short-Term Memory (LSTM) networks and demonstrates their ability to solve complex sequence learning tasks with long time lags, outperforming traditional recurrent neural networks and other methods.

Andrew Ang et al. (2005)

The paper establishes two key findings: aggregate volatility risk is priced with a negative premium (investors pay for insurance), and stocks with high idiosyncratic volatility exhibit anomalously low returns (-1.31% monthly alpha) that cannot be explained by standard factors.

Factor Models of Asset Returns

Gregory Connor and Robert Korajczyk (2009) · 17 citations

This paper surveys how factor models decompose asset returns into pervasive and idiosyncratic components, and explains major model classes (statistical, macro, characteristics), estimation methods, and how to choose the number of factors.

Aapo Hyvärinen et al. (2010) — Journal of Machine Learning Research · 440 citations

This paper introduces a novel method for estimating Structural Vector Autoregression (SVAR) models by leveraging the non-Gaussianity of disturbance variables, overcoming identifiability issues common in traditional SVAR estimation.

Andrew J. Patton (2012) — Journal of Multivariate Analysis

A survey of how copulas let economists model time-series marginals (means/volatility/tails) separately from cross-asset or serial dependence—covering estimation, inference, goodness-of-fit tests, and key finance applications.

Juan Peralta Donate et al. (2013) — Neurocomputing

The paper proposes evolving an ANN ensemble with a time-ordered, weighted n-fold cross-validation fitness to improve short/medium time-series forecasts, achieving lower SMAPE than standard (unweighted) CV ensembles and a single holdout-evolved ANN, and matching Holt–Winters on average.

Balaji Lakshminarayanan et al. (2017) — Curran Associates, Inc. · 7168 citations

This paper proposes Deep Ensembles, a simple and scalable non-Bayesian method combining multiple neural networks, proper scoring rules, and adversarial training to provide high-quality, well-calibrated predictive uncertainty estimates that are robust to out-of-distribution data, outperforming or matching approximate Bayesian methods.

Christoph Bergmeir et al. (2018) — Computational Statistics & Data Analysis

The paper shows that standard random K-fold cross-validation can be valid for autoregressive time-series forecasting—if (and largely only if) the model’s forecast errors are serially uncorrelated—often making it better than single holdout out-of-sample evaluation.

Markus Pelger (2019) — Journal of Econometrics

This paper develops a statistical theory and practical estimators for large-dimensional factor models using high-frequency financial data, enabling the separation of continuous and jump risk factors and providing a robust method for determining the number of factors.

Jakob Runge et al. (2019) — Science Advances

The paper introduces PCMCI, a two-stage causal discovery method for high-dimensional, autocorrelated, nonlinear time series that improves true-link detection while controlling false positives compared with Granger/FullCI, PC, and Lasso.

Nick Baltas and Robert Kosowski (2019)

Optimizes Time-Series Momentum (TSMOM) by reducing turnover ~36% via efficient volatility/trend estimation and improving post-crisis performance through dynamic correlation-based leverage.

Omer Berat Sezer et al. (2019) — arXiv:1911.13288 [cs, q-fin, stat]

This paper provides a comprehensive systematic literature review of Deep Learning (DL) models applied to financial time series forecasting between 2005 and 2019, highlighting model choices, performance against traditional Machine Learning (ML), and future research directions.

Michael Pinelis and David Ruppert (2022) — The Journal of Finance and Data Science

The paper shows that a utility-maximizing market-timing rule that combines machine-learned forecasts of next-month equity premium and volatility ("reward-risk timing") materially improves Sharpe ratio, drawdowns, and investor utility versus buy-and-hold.

Andrew Ang (2023) — The Journal of Portfolio Management

A regime-based analysis of five major equity style factors showing that Value and Momentum have structurally deteriorated in the 21st century while Quality and Size have improved, decomposing returns into long-term trends and short-term cycles.

Michele Leonardo Bianchi et al. (2023) — The Journal of Portfolio Management · 8 citations

This tutorial-style paper explains why Gaussian (normal) return models underestimate extreme losses and surveys univariate and multivariate fat-tailed alternatives (especially Lévy-process-based) that are more suitable for pricing, risk, and portfolio construction.

Kashif Rasul et al. (2024)

Lag-Llama is a decoder-only transformer foundation model for univariate probabilistic time series forecasting, pretrained on a large, diverse corpus, demonstrating strong zero-shot generalization and state-of-the-art performance after fine-tuning on unseen datasets.

Agathe Sadeghi et al. (2024)

The paper proposes CD-NOTS, a constraint-based causal discovery algorithm tailored to nonstationary time series that explicitly models lagged dependencies, and shows in simulations and finance case studies that it can recover more reliable causal graphs than common baselines when stationarity/linearity assumptions fail.

Ronald Hua and Clare Sun (2024)

The paper proposes a practical framework to quantify how much historical factor information should be trusted out-of-sample under market nonstationarity, showing that optimal “shrinkage toward 1/N” depends strongly on market efficiency.

Miquel Noguer I Alonso (2024)

Conformal prediction is a statistical framework that generates reliable, distribution-free prediction intervals for financial data, enhancing risk management, portfolio optimization, and algorithmic trading, as demonstrated in a banking stress test case study.

Abhimanyu Das et al. (2024)

This paper introduces TimesFM, a decoder-only foundation model for time-series forecasting that achieves near state-of-the-art zero-shot accuracy across diverse datasets by pretraining on a large corpus of real-world and synthetic time-series data.

Zhen Zeng et al. (2024)

The paper turns time series into time–frequency spectrogram images (plus a row encoding the raw series) and shows a Vision Transformer can forecast better than line-plot images and standard baselines across synthetic, weather, and equity-price datasets.

Albert Gu and Tri Dao (2024) · 6280 citations

Mamba introduces a novel selective state space model (SSM) architecture that achieves Transformer-level performance on diverse modalities with linear scaling in sequence length, enabling faster inference and longer context processing.

Kevin J. DiCiurcio et al. (2024) — The Journal of Portfolio Management

A two-stage framework that first uses unsupervised learning to identify market risk regimes (Normal, Correction, Bear) and then uses supervised learning to predict winning equity factors within those regimes, outperforming static and business-cycle approaches.

Gerald Woo et al. (2024)

This paper introduces MOIRAI, a Transformer-based large time series model trained on a new 27B-observation dataset (LOTSA) to achieve competitive zero-shot forecasting across diverse time series, addressing challenges like varying frequencies, arbitrary variates, and diverse distributions.

Mononito Goswami et al. (2024)

MOMENT is an open-source family of large transformer models pre-trained on a diverse collection of time series data, demonstrating strong performance across various time series tasks with minimal fine-tuning, especially in limited supervision settings.

Xu Liu et al. (2024) · 75 citations

This paper introduces MOIRAI-MOE, a time series foundation model that uses sparse Mixture of Experts (MoE) to achieve data-driven, token-level specialization, significantly outperforming existing models in zero-shot forecasting by overcoming limitations of human-imposed frequency-level specialization.

Matt Lutey (2025) — The Journal of Portfolio Management

The paper is an instructional tutorial showing how to convert financial price time series into images (e.g., Gramian Angular Fields and Recurrence Plots) so CNNs can learn chart-like patterns for forecasting, screening, and risk monitoring.

Yong Liu et al. (2025)

This paper introduces Timer-XL, a causal, decoder-only Transformer that generalizes next token prediction to multivariate time series, achieving state-of-the-art unified and zero-shot forecasting performance by effectively handling excessively long contexts and complex inter-series dependencies.

Miquel Noguer i Alonso (2025)

This paper introduces a comprehensive framework for Financial Information Theory, applying information-theoretic concepts to financial time series and demonstrating their empirical usefulness for regime detection, market efficiency testing, and portfolio construction using S&P 500 data from 2000-2025.

Kun Feng et al. (2025)

KAIROS is a novel time series foundation model that uses adaptive tokenization and instance-specific positional embeddings to achieve superior zero-shot forecasting performance with fewer parameters on diverse and heterogeneous time series data.

Xiaoming Shi et al. (2025)

TIME-MOE introduces a scalable, sparse Mixture-of-Experts (MoE) Transformer architecture for time series forecasting, pre-trained on a 300 billion time point dataset, achieving state-of-the-art accuracy and efficiency by scaling to 2.4 billion parameters while reducing inference costs.

Dragos Gorduza et al. (2025) — EPJ Data Science

This paper introduces a novel approach to forecast stock market volatility by using the edge reconstruction accuracy of a graph auto-encoder (GAE) to measure the homogeneity of connections between stocks, showing that lower homogeneity correlates with higher volatility and improves out-of-sample volatility forecasting.

Matthieu Boileau et al. (2025)

The paper shows that basic time-series interpretation (trend, noise, extrema location) can be distilled from a large multimodal model into very small instruction-tuned LMs, enabling lightweight, natural-language explanations of temporal patterns.

Eghbal Rahimikia et al. (2025) · 1 citations

Time Series Foundation Models (TSFMs) show promise for financial forecasting but require domain-specific pre-training on large financial datasets to outperform traditional ML models and achieve economic gains.

Selin Yilmaz and Emre Sefer (2026) — The Journal of Finance and Data Science

This paper explores the application of time-series-based deep learning techniques, including LSTM, Informer, Autoformer, iTransformer, Scaleformer, and Chronos, to pairs trading strategies, demonstrating their potential to outperform traditional statistical arbitrage methods by more accurately predicting residual fluctuations between cointegrated asset pairs.

Philip Hans Franses — International Journal of Forecasting

The paper argues that MASE (Mean Absolute Scaled Error) is a forecast accuracy measure that “plays nicely” with Diebold–Mariano equal-accuracy tests because it preserves the moment conditions needed for asymptotic normality, unlike percentage-error based measures.

Xiaofan Zhou et al. — ACM Comput. Surv.

This survey provides a comprehensive, data-centric overview of Conformal Prediction (CP), a distribution-free uncertainty quantification framework, detailing its foundational concepts, recent advancements, and applications across diverse data modalities like structured, unstructured, and dynamic data, while also discussing open challenges.

Min B. Shrestha and Guna R. Bhatta — The Journal of Finance and Data Science

The paper provides a step-by-step framework for choosing the right econometric method for time-series data (unit roots, cointegration, ECM/ARDL), and shows on Nepal inflation data how naïve OLS/VAR can look “great” yet be spurious.

Zhe Li et al. — Association for Computing Machinery

This paper introduces TSFM-Bench, a comprehensive and unified benchmark for Time Series Foundation Models (TSFMs), evaluating their performance across diverse datasets and forecasting scenarios (zero-shot, few-shot, full-shot) using standardized protocols.