Neural Realized Measures: Forecast-Supervised Intraday Aggregation for Volatility Forecasting
In briefNeural Realized Measures: Learning Intraday Aggregation for Volatility Forecasting
SummaryStandard volatility forecasting compresses intraday observations into a daily realized measure using a fixed rule before forecasting. This paper proposes Neural Realized Measures (NRM), which keeps the established intraday contributions but replaces uniform weights with positive, mean-normalized context-dependent weights learned from the forecasting objective. NRM nests the classical raw realized measure as a special case, and in Heston simulations and five-minute cryptocurrency data it improves forecast accuracy in most comparisons.