RAVEN: A Regime-Aware Variable-context Expert Network for Financial Time Series Forecasting

A new financial forecasting model called RAVEN adapts its look-back window to shifting market regimes, addressing a structural weakness in fixed-context time series architectures, according to research published on arXiv [1]. The Regime-Aware Variable-context Expert Network, or RAVEN, is a Mixture-of-Experts framework that determines the temporal context for each input sample rather than relying on a fixed horizon [1]. Financial log-returns are non-stationary, carry low signal-to-noise ratios, and are governed by regime-dependent temporal dependencies, the authors note [2]. Standard state-of-the-art time series models apply a fixed context window, which is mismatched to the time-varying optimal look-back of non-stationary price processes [2]. RAVEN instead constructs a hierarchy of nested contiguous windows whose lengths are determined by the data itself [1]. The model scores patches by learned importance in reverse chronological order and applies a Cumulative Importance Thresholding mechanism to derive nested prefix windows, each routed to a scale-specialized expert [2]. A Global Compressed Representation branch runs in parallel over the full context, preserving global temporal coherence that local experts cannot guarantee [2]. Because the nested routing induces structured overlap among expert inputs, the authors introduce a Correlation-Aware Weighting to align variable-length expert outputs and penalize pairwise cosine similarity prior to aggregation [2]. On cumulative log-return prediction for the HS300 and S&P500 indices, RAVEN improved Pearson correlation by 9.2% and 20.2%, respectively [1]. On a fund sales forecasting task, the model reduced mean squared error by 18.2% [1]. The paper further reports that RAVEN achieved the best results in 14 of 16 metrics on four PEMS traffic benchmarks, suggesting the adaptive-context design generalizes beyond financial data [2]. The work addresses a long-standing challenge in financial machine learning: the non-stationary nature of price processes means the relevance of historical observations changes over time, yet most deep learning architectures treat all past data within a fixed window as equally informative [2]. By learning which patches matter and routing them to specialized experts, RAVEN tailors its receptive field per sample, a departure from one-size-fits-all sequence models [1]. The research appears as a preprint and has not yet been peer-reviewed [1].