This study evaluates multiple forecasting models, ranging from HAR and GARCH to Tree-based and Neural architectures, across 14 Global Equity Indices and Horizons form 1 day to 100 Trading days within a strictly chronological and capacity-controlled framework, indicating that for strongly dependent time series, nominal sample size is a misleading measure of learnability, and model capacity must be constrained relative to effective information rather than observation count.
Financial volatility is regime dependent, yet incorporating regime information into neural networks can also destabilize training. This paper asks where such information should enter a neural cross-sectional volatility forecasting model. We study five-day realized-volatility forecasts for 1,027 U.S. equities using a rolling walk-forward evaluation framework in which information, model capacity, hyperparameter tuning, and random seeds are matched across architectures. We propose RG-ResMoE, a regime-gated residual mixture-of-experts architecture in which regime information is used only for expert routing rather than for direct forecasting. The base predictor models volatility from stock features, while a gating network uses regime state variables to route residual corrections. RG-ResMoE consistently outperforms a capacity-matched MLP in both forecasting accuracy and training stability in the main U.S. study. Similar gains are observed on an independent Japanese panel. The integration pathway is decisive: appending the same regime variables directly to the forecasting input degrades both predictive performance and training stability, whereas restricting them to the routing gate improves accuracy and Value-at-Risk calibration. Hard routing consistently underperforms soft routing. The results suggest that, in compact neural volatility forecasting models, the primary value of mixture-of-experts models lies less in increasing model capacity than in controlling how nonstationary regime information influences prediction.
SHAP analysis reveals that trading activity, lagged liquidity, and market uncertainty are the main determinants of liquidity forecasts, and the findings highlight the complementary role of explainable machine learning in empirical finance.
Veni Arakelia, G. Caporale, Mirto M Gasparinatou et al.· CESifo working papers· 0 citations
Forward-looking volatility forecasts are central inputs to derivatives pricing, market making, risk management, and volatility-linked trading strategies, with ARCH and GARCH models serving as the canonical workhorses. Such models are natural in standard asset markets, where prices are positive-valued stochastic processes and volatility is typically inferred from return dynamics. Prediction markets have a different structure: prices are bounded probabilities, payoffs are binary, and contracts resolve at known deadlines. We develop and estimate a volatility model tailored to binary prediction markets. The model combines two economic mechanisms: a Wright-Fisher deadline-resolution component, capturing how remaining binary uncertainty is forced to resolve over time, and a Glosten-Milgrom order-flow component, capturing volatility from informed trading as reflected in spreads and volume. Using a large panel of Kalshi contracts, we show that these structural variables carry substantial forecasting power. Plain ARCH/GARCH benchmarks are dominated by structural specifications; combining the structural model with residual GARCH dynamics gives the best overall forecasts. The model also provides an interpretable measurement framework: volatility is highest near fifty-fifty prices, rises near resolution, and varies across categories with the timing and discreteness of information arrival. Economics contracts are closer to smooth deadline-resolution dynamics, while sports contracts exhibit more event-concentrated, jump-like behavior. Across major categories, category-specific fitting does not systematically improve out-of-sample performance, suggesting that the structural specification transfers beyond the pooled headline result.
Weiye Xi, C. Moallemi, Mallesh M. Pai et al.· 0 citations
Volatility forecasts play an important role in financial markets. They are used for setting position limits, determining margin requirements, and guiding short-term trading decisions. Financial volatility is known to exhibit persistence and regime dependence, which makes short-horizon forecasting challenging. In addition to historical volatility measures, implied volatility indicators such as the Volatility Index (VIX) are often used as forward-looking proxies for market risk expectations. However, it is not clear whether the predictive usefulness of the VIX is consistent across different neural network architectures. This study investigates the role of the VIX in one-day-ahead realized volatility forecasting and examines whether its contribution varies across alternative neural network models. The empirical analysis focuses on major global equity markets, including the S&P 500, DAX, FTSE 100, Nikkei 225, and Hang Seng Index. Using market data from these indices, the study evaluates the forecasting performance of artificial neural networks (ANN), long short-term memory networks (LSTM), and gated recurrent unit (GRU) models. All models are estimated within a rolling out-of-sample framework and are tested with and without the VIX variable. Forecast accuracy is evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Quasi-Likelihood (QLIKE) loss functions, and Diebold–Mariano tests. The results show that memory-based models outperform feedforward ANN models, while the contribution of the VIX varies across models and markets.
Orhan Özaydın· Journal of Intelligent Decis...· 0 citations
This paper compares direct-yield and factor-based approaches to U.S. Treasury yield curve forecasting using a common high-dimensional macroeconomic information set. Forecasts are evaluated on monthly zero-coupon yields over the 2015-2025 out-of-sample period. Gains over the random walk are concentrated at short maturities and in slope forecasts, and decline with the forecast horizon. Direct-yield models perform best for slope forecasts and are relatively stronger at short horizons, while factor-based models become more competitive at longer horizons. Macroeconomic predictors provide clear incremental predictive power, strongest for slope-related movements. A trading simulation reinforces that macro-augmented models perform best in slope trades. The simulation also highlights a gap between statistical and economic performance, as the random walk is a strong benchmark under statistical loss but performs poorly as a trading signal.
This paper utilises high‐frequency data from the Chinese stock market and a panel of individual stocks to compare the forecasting performance of machine learning with popular econometric models across different periods. For short‐term volatility forecasting, most machine learning models outperform econometric models with limited explanatory variables, though they do not exhibit a significant advantage over the econometric model incorporating all features. For medium‐ and long‐term volatility forecasting, Light gradient boosting machine (LGBM) in machine learning substantially dominates econometric models. We also explore a simple‐to‐implement forecast combination method that leverages the best machine learning model and the best econometric model to explore if model averaging leads to any improvement. Our findings indicate that over a longer forecasting horizon, this method achieves the best performance among all forecast combinations and dominates the best econometric model.
