A volatility net that repeatedly fails to train is worse than one that settles for mediocre forecasts. From a trading desk, the headline result in Ye and Borde is 0 of 30. Across thirty random seeds on the US equity panel, every run finishes below the collapse threshold fixed in advance, a mean test QLIKE above 2.0. A standard mixture-of-experts collapses in 24 of 30 under the same tuning protocol, using identical features and matched capacity. A desk can use that stability gap. The paper's ablations establish it clearly.

What they built

The model predicts each stock's annualized five-day forward realized volatility. Performance is the daily cross-sectional Spearman correlation between each forecast and its outcome. The US panel contains 1,027 common stocks from Yahoo Finance and runs from December 2015 to November 2025. Replication uses a 1,552-name TSE Prime panel processed in the same way.

Sixteen stock-level inputs feed the model: trailing 5, 20 and 60 day realized vol, cumulative 5 and 20 day returns, 14-day RSI, and the last ten daily log returns. Evaluation starts with a 504-day development window, split 85% train and 15% validation. A 63-day test window follows, then the process advances 63 days and repeats thirty times. The resulting sequence spans April 2018 to October 2025, covering 1,890 test days and roughly 1.9 million out-of-sample forecasts.

RG-ResMoE, their regime-gated residual mixture-of-experts, trains in two stages. First, a two-hidden-layer GELU block learns from stock features alone and is then frozen. Four small experts learn corrections to its forecast. Zero initialization in each expert's output layer makes the starting model an exact copy of the base. A softmax gate mixes those corrections.

Only the gate receives the regime state. Its inputs are 20-day rolling volatility of the equal-weighted market return and 20-day rolling volatility of the residual return, using beta from the preceding 120 days. Two variables. The loss penalizes the aggregate correction's size and adds a load-balancing term that pushes routing weights toward 1/K.

The paper's claim rests on the entry point for regime information. When the same two variables are appended to the forecasting input, MLP-L loses 0.0043 of IC, falling from 0.5421 to 0.5378. Its collapsed seeds rise from 3 to 30 of 30. Ridge loses 0.0023. MLP-S gives up 0.0011, while collapses jump from 2 to 22. Sending the variables through the gate moves IC from 0.5466 to 0.5469 and removes the one remaining collapsed run.

Where the 0.0048 comes from

MLP-L is the capacity-matched baseline, H=44, sized against the complete RG-ResMoE. The model leads every metric on the US panel. IC is 0.5469 +/- 0.0012 against 0.5421 +/- 0.0033. RMSE is 0.2304 vs 0.2320, while out-of-sample R-squared is 0.292 vs 0.282. ICIR comes in at 6.14 vs 6.08. QLIKE is 0.735 vs 2.23.

The 0.0048 improvement equals about one and a half times MLP-L's own across-seed standard deviation. Significance is based on paired Diebold-Mariano tests using about 1,900 daily seed-averaged differences, with Newey-West errors at four lags. Four lags fits the overlapping five-day targets.

Now place the gate next to those results.

Across the pathway comparison on the same US panel, the complete architecture adds +0.0048 over MLP-L. The regime pathway supplies +0.0004. Most of the improvement comes from the frozen base and zero-initialized residual head. Randomly initialize the residual layer and IC drops 0.0039, while collapses climb from 0 to 13 of 30. Remove the frozen base entirely, producing the standard MoE, and the loss is 0.0056 with 24 collapses.

Section 6.4 identifies the residual architecture as the principal contribution, although the abstract leads with the pathway. The authors give a direct defence: input concatenation damages accuracy and stability, while routing through the gate improves accuracy and VaR calibration, making the pathway decisive. Its defensive effect is large, with 30 of 30 collapses and -0.0043 IC for MLP-L(+z). Its positive contribution is +0.0004. The gate offers a safe home for the variables rather than much additional edge.

Consider the scale of 0.0004. Twenty-day persistence records IC of 0.5018 on this panel, compared with 0.5469 for RG-ResMoE. The entire distance between a rule that fits in one line and the best model here is 0.0451. Regime routing accounts for nine tenths of one percent of that distance. I would skip the gating network. The frozen base and zero-init experts are worth about a tenth of the same distance, and they bring the stability too.

The frozen base drives that stability. The paper also credits the residual design and frozen base as the primary sources of training stability. Standard MoE, capacity-matched to RG-ResMoE, delivers out-of-sample R-squared of 0.259 on the US panel. Plain Ridge reaches 0.273. Pooled HAR records QLIKE of 0.749 against RG-ResMoE's 0.735, a two percent difference. HAR is deterministic, removing seed risk from the decision. For variance calibration, HAR sits closer than the IC table implies.

Japan preserves the ordering. On that panel, RG-ResMoE posts IC of 0.4858 +/- 0.0024 against 0.4815 +/- 0.0036 for MLP-L, an improvement of 0.0043 at p<1e-4. QLIKE is 0.80 against 11.6. Collapses are 1 of 30 against 6 of 30. Appending the two regime variables to MLP-L produces collapses in 28 of 30.

Can you size risk with it?

The deployment argument is strongest in the VaR results. Their absolute rejection rates still keep me from sizing on the model. RG-ResMoE has the lowest Kupiec coverage rejection rate in every slice, and its margin over both MLP baselines is significant in five of six. At 5% VaR, rejection covers 26.7% of tickers, versus 30.3% for MLP-L and 32.7% for MLP-S. During high-vol periods, the figures are 11.9% against 14.8% and 16.9%. In the 1% low-vol slice, RG-ResMoE ties MLP-L at 5.8%, without significance, as the authors report.

The level matters alongside the ranking. Even the best model misses unconditional coverage for 26.7% of names at 5% VaR. Failures cluster in calm periods, reaching 22.9% in low-vol conditions against 11.9% in high-vol conditions. Different day counts give those slices different test power, leaving the contrast unfit for a clean comparison.

Stress periods provide the result that could change my sizing. Against MLP-L, the IC advantage reaches +0.0207 in the top market-vol decile, covering 189 days. From February to June 2020, it rises to +0.0322 across 104 days, compared with +0.0048 over the full sample. Regime-flip windows cover 826 days and produce +0.0031, below the full-sample improvement. The paper reports this result and leaves it there. My reading is that it cuts against a regime-switching story.

Survivorship. Still no P&L

The universe includes only stocks listed at the sample's end and requires at least six years of history. Current S&P 1500 constituents account for 85%, with membership observable only ex post. The authors disclose the restriction. Since it applies to every architecture in the comparison, the ordering probably remains. Absolute IC and especially the VaR tail come from a panel stripped of delisted names.

There is no portfolio here. No returns, no Sharpe, no turnover, no cost assumption.

The full-sample improvement is 0.0048 across 1,890 days. On the 189 highest-volatility days, it reaches 0.0207. Either figure must survive conversion into position sizes before it becomes money. We have previously covered vol papers that are precise about estimation and silent about the trade (our note on EGARCH asymmetry). We are rebuilding this now. The question is whether the forecast changes sizing enough to appear in realized drawdown.

Median-seed comparisons would move me away from the sceptical reading. MLP-L's QLIKE of 2.23 has a standard deviation of 5.15, indicating that a few catastrophic seeds drive the figure. The per-seed scatter makes the distribution visible, although I did not find a median-seed table. If the median MLP-L seed matches RG-ResMoE on QLIKE, the claim contracts to tail insurance against training failure. With 24 of 30 collapses for standard MoE and 30 of 30 for MLP-L when regime inputs are appended, that insurance remains worth having.