Answers are grounded in this paper and cite their sources.
Weekly ETF return forecasting is constrained by weak signals, correlated predictors, and point-in-time information availability. This study examines how feature architecture and regularisation affect forecast loss and availability for three focal ETFs and 113 mainland-listed domestic equity ETFs. Nine methods are compared at a common weekly horizon under four ex ante architectures: return state, ETF trading liquidity, market state, and a combined pool. Regularised and tree estimators use 24 time-ordered validation origins, while ARIMA-family orders are selected by AICc on a fixed 26-week recalibration calendar. On matched dates, trading liquidity produces lower forecast loss than the combined pool in 16 of 18 focal and 457 of 675 eligible external comparisons. Under lag-zero variance estimation and Benjamini–Hochberg adjustment, 48 external differences remain significant, all favouring trading liquidity. HAC–BH retains 30 in that direction and one Random Forest case favouring the combined pool, while lag-zero Benjamini–Yekutieli retains 14 favouring trading liquidity. No complex-model improvement over the Historical Mean survives multiplicity adjustment. Ridge provides complete focal-core and external forecast coverage, whereas LASSO–ARIMAX can be unavailable after zero-regressor screening. These findings highlight the importance of information timing, predictor organisation, and forecast availability. This study examines how real-time information availability and predictor organisation affect weekly Chinese ETF return forecasts. Using three focal ETFs and a broader sample of 113 funds, it evaluates nine forecasting methods and four feature architectures under strict point-in-time validation rules. Results show that focused trading-liquidity predictors frequently reduce forecast loss compared to a combined predictor pool, though complex models rarely outperform the simple Historical Mean once statistical testing accounts for multiple comparisons. For researchers and quantitative practitioners, these findings underscore the practical importance of strict information cutoffs, compact feature sets, and evaluating whether models reliably generate usable predictions alongside traditional accuracy metrics.
No takes yet. Share an insight, caveat, or question.
Yuan Lu (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: