Regime-Gated Residual Mixture-of-Experts for Cross-Sectional Volatility Forecasting Junyi Ye, Gargi Vijay Borde · ICAIF '26 · paper PDF · doi:10.1145/3858063.3858798
This repo contains the code and data behind the paper. It forecasts five-day stock volatility for over 1,000 U.S. equities, and shows that a mixture-of-experts model does best when market-regime information is used only to route between experts, never fed into the forecast itself.
Financial markets move through regimes: calm periods, crises, recoveries. A growing line of work uses mixture-of-experts (MoE) architectures to handle this, on the idea that different experts can specialize in different market conditions. These models often beat plain baselines. But it's rarely clear why.
A typical MoE setup changes three things at once relative to its baseline. It exposes the model to regime information it didn't have before. It adds parameters, usually a lot of them, since an MoE baseline is commonly compared against a much smaller plain MLP rather than one with a matching parameter count. And it introduces a new mechanism: routing between specialized sub-models. Any of the three could be driving the improvement, but most papers compare against an MLP with far fewer parameters, so capacity and architecture are never actually separated.
This paper isolates that question. Every model we compare has a matched parameter budget, the same regime information available to it, and the same tuning/evaluation protocol. The only thing we vary is how, and whether, regime information reaches the model. That lets us ask cleanly: does the gain come from having the information, from having more capacity, or from the gating/expert architecture itself?
The answer is the architecture, not capacity. Against a plain MLP with the same parameter count as RG-ResMoE, giving that MLP the same regime variables at its input still hurts it, while routing those same variables through an MoE gate helps. Same information, same capacity, opposite outcome. The benefit comes from where the information enters the model, not from how much of it there is, and not from having more parameters.
RG-ResMoE has two parts:
- A frozen base network predicts volatility from a stock's own price history.
- A small team of correction experts learns to adjust that forecast. A gate decides how much weight each expert gets, and it's the only part of the model allowed to see market-regime signals (overall market volatility, and a stock's own idiosyncratic volatility).
Keeping regime information out of the forecast and only in the gate turns out to matter a lot: it beats an equally-sized MLP on accuracy, risk calibration, and training stability, and the same result holds on an independent Japanese stock panel.
pip install -r requirements.txt
bash reproduce.sh # builds the data panel, trains every model, aggregates resultsThat reproduces the paper's main U.S. results end to end: Table 3 (headline forecasting results), Table 4 (the regime-information pathway comparison), Table 7 (VaR coverage) and Table 8 (stress slices). It needs a CUDA GPU for training (one model/seed per GPU); each step skips work that's already done, so it's safe to stop and resume.
Want more than the headline table?
RUN_ABLATIONS=1 bash reproduce.sh # + routing/design ablations (Tables 5-6)
bash reproduce_jp.sh # + the Japan replication (Table 9)Budget roughly, at 30 seeds: the default U.S. run is ~12 GB of per-seed
prediction files and a few hours of GPU time (plus ~10 min of CPU for the
per-stock GARCH fits); adding RUN_ABLATIONS=1 takes it to ~32 GB, and the
Japan replication adds ~11 GB. The feature panels themselves are ~1.5 GB.
Everything lands under data/processed/ and results/, both gitignored.
src/ Pipeline code: build the data panel, train each model, score results
data/raw/ Daily U.S. stock prices (1,029 tickers), exactly what the paper used
data/raw_jp/ Daily Japanese stock prices (1,552 tickers), for the replication study
results/reference/ The paper's own output, so you can check your run against it
Everything else (data/processed/, results/predictions/,
results/summary/) gets created as you run the pipeline.
src/train_rg_resmoe.py is the main model. A handful of environment
variables switch it into the other configurations reported in the paper:
expert count, routing style, whether the gate sees regime info. The
baselines and routing variants it's compared against live in
src/train_mlp.py, src/train_standard_moe.py, src/train_rule_gate_moe.py,
src/train_ridge.py, and src/train_classical_baselines.py (Persistence,
HAR, GARCH).
If you're trying to match a specific table or number in the paper, open
reproduce.sh instead. It lists every model/flag combination as a single
runnable command, in the order the tables need them.
- The shipped CSVs are the exact data used in the paper. Yahoo Finance revises its history over time, so a fresh download won't match bit-for-bit.
- Every neural model is trained with 30 random seeds; significance tests average results across seeds first, then compare day by day (Diebold-Mariano, Newey-West lag 4).
- The Japan replication reuses the hyperparameters selected on the U.S. data: the paper found the same configuration worked best on both.
results/reference/holds the paper's 30-seed numbers, with row labels kept in step with the current scripts so a fresh run lines up directly. One row is absent:collapse_rate_by_model.csvhas noRG-ResMoE-K6, because that variant was added after this reference was generated.
If you find this repo useful, please consider giving it a star ⭐ and citing the paper:
@inproceedings{ye2026rgresmoe,
title = {Regime-Gated Residual Mixture-of-Experts for Cross-Sectional Volatility Forecasting},
author = {Ye, Junyi and Borde, Gargi Vijay},
booktitle = {Proceedings of the 7th ACM International Conference on AI in Finance (ICAIF '26)},
year = {2026},
address = {Milan, Italy},
publisher = {ACM},
doi = {10.1145/3858063.3858798}
}MIT (see LICENSE). Raw price data derive from Yahoo Finance and are provided for research reproducibility only. The paper PDF is CC BY 4.0.