RAVEN Improves Financial Time Series Forecasting with Adaptive Context

Financial time series forecasting presents unique challenges due to the non-stationary nature of log-returns, low signal-to-noise ratios, and regime-dependent temporal dependencies. A key limitation of current state-of-the-art models is their reliance on fixed context windows, which are ill-suited for the time-varying optimal look-back periods inherent in financial price processes. To address this, researchers propose RAVEN (Regime-Aware Variable-context Expert Network), a Mixture-of-Experts framework that dynamically adjusts the temporal context for each input. Instead of a static look-back horizon, RAVEN constructs a hierarchy of nested contiguous windows whose lengths are determined by the data itself. It scores data patches by learned importance in reverse chronological order and applies a Cumulative Importance Thresholding (CIT) mechanism to derive these nested windows, each routed to a specialized expert. RAVEN also includes a Global Compressed Representation (GCR) branch that processes the full context to maintain global temporal coherence, complementing the local experts. Furthermore, a Correlation-Aware Weighting (CAW) mechanism aligns the variable-length expert outputs and penalizes cosine similarity before aggregation, accounting for the structured overlap in expert inputs. Experiments on cumulative log-return prediction for major indices (HS300, S&P500) and fund sales forecasting show RAVEN achieving state-of-the-art results, with significant improvements in Pearson correlation and reduced Mean Squared Error, also performing well on traffic benchmarks.