WHICH MODEL FAMILIES PAY? ECONOMETRIC, GRADIENT BOOSTING AND RECURRENT NEURAL NETWORK VOLATILITY FORECASTS IN ACTIVE TRADING STRATEGIES ON THE RUSSIAN STOCK MARKET

Abstract

This paper asks which families of volatility forecasting models create economic value in active trading, and through which integration channel that value is transmitted. Seven models drawn from four families — econometric (GJR-GARCH, HAR-J), gradient boosting (XGBoost, LightGBM), recurrent neural networks (LSTM, GRU) and a hybrid combining HAR-J with boosting — are compared on the realized volatility of 17 liquid Moscow Exchange stocks computed from 10-minute returns over 2014–2026, using a feature space of 234 variables and QLIKE as the loss function. Two findings follow. First, the machine learning advantage is neither uniform across families nor stable across horizons: gradient boosting improves on the HAR-J benchmark by 5–9% at the one-day horizon but loses to it at the five-day horizon under rolling re-estimation, whereas recurrent networks are 22–31% worse than HAR-J at the one-day horizon and dominate it at no horizon; only the hybrid ranks at or near the top throughout. Second, the translation of accuracy into trading performance is governed by the integration channel rather than by the size of the accuracy gain: a 28% reduction in QLIKE raises the Sharpe ratio of a six-strategy portfolio by +0.79 and +0.85 under regime filtering and volatility targeting (p < 0.001), but by an insignificant +0.04 when the forecast only scales the parameters of an individual trade. Model selection therefore pays where the forecast governs the entry decision and the position size, and not otherwise. Statistical accuracy is a valid proxy for economic value in this setting, but only conditionally on the channel through which the forecast reaches the trading decision.