Predicting CSI 800 Sector Index Returns in China: A Multifactor Statistical and Regularized Learning Approach

Abstract

Industry allocation is an important component of asset allocation and a major source of excess returns in active investing. This study investigates the predictability of sector index returns in China using a multifactor framework. Using monthly data from January 2019 to June 2026 for eleven CSI 800 first-level sector indices, we construct 11 factors including indicator (CPI, PPI, PMI, M2), interest rates (Shibor), technical indicators (momentum and volatility), and a market factor. We model each industry separately and estimate by using two linear models: ordinary least squares (OLS) and Lasso regularized regression, with the period from January 2025 to June 2026 as an out-of-sample test window. A one-shot training scheme (fixed sample of January 2019 to December 2024) yields negative out-of-sample R² for both models (−0.72 for OLS and −0.12 for Lasso), which was normal in monthly financial return prediction. To improve the forecasts, we adopt rolling window re-estimation: at each month's end, the models are refitted on the most recent 72, 60, and 48 months of data. Rolling-window estimation is associated with improved numerical accuracy and strategy returns: the 72-month rolling Lasso raises the average out-of-sample R² to −0.07 and delivers a Top-3 industry rotation strategy with a cumulative return of 65.45%, an excess return of 48.03 percentage points over the equal-weight benchmark; the 72-month rolling OLS earns 37.46% (excess +20.04%). Since both models produce negative R² for out-of-sample prediction, we further investigate their ranking power using Spearman rank information coefficient (Rank IC) calculated on a monthly cross-section basis: using the 72-month rolling window approach, the mean Rank IC equals +0.215 (t = 2.96) for Lasso, compared to just +0.057 (t = 0.60) for OLS. Thus, the ranking power of Lasso strategy provides support to the presence of excess returns, while the ranking power of OLS strategy is not statistically different from zero. These results suggest that rolling-window estimation can help models adapt to changing market conditions. Compared with OLS, Lasso shows stronger out-of-sample ranking performance, which is consistent with the excess returns generated by the corresponding industry rotation strategy.

Read full paper (PDF) →