Estimating SET100 realized beta with long short-term memory model

Abstract

This study examines whether a Long Short-Term Memory (LSTM) network, trained on ten firm-level and macroeconomic predictors, can forecast the one-year-ahead realized beta of SET100 stocks more accurately than the one-year daily and five-year monthly rolling regressions that practitioners rely on, using monthly out-of-sample forecasts from 2015 to 2024 evaluated on a value-weighted basis. The result suggests that the LSTM forecasts beta more accurately than the five-year monthly regression in every size group and in both COVID-19 regimes. However, it does not dominate the one-year daily regression. In fact, the two models are statistically equal on the value-weighted aggregate, and this parity holds both before and after the COVID-19 break, contradicting a common expectation that a flexible deep-learning model should beat simple regressions across the board. A possible explanation for this finding is that a one-year daily regression already estimates the betas of large and liquid stocks with little bias, so little room is left for a richer model to add value among the stocks that dominate the value-weighted measure. The network instead earns its advantage among small and mid-sized stocks, significantly so in the mid group, and its forecasts depend mainly on the recent historical betas rather than on the wider predictor set. Finally, these gains are observed in some, but not all, segments of the market, as the advantage among smaller stocks is clearest before the pandemic, so the additional cost of an LSTM is justified mainly for the small and mid-cap segment.