Bastola and Prushansky's NVDA results favor a simple trade: fade an implied-vol spike and expect small losses on most days. The neural network adds nothing to that rule in their results. Almost all the case for taking it rests on about five sessions out of 281.

The forecast being tested

The authors build a split-adjusted NVDA panel from May 2018 to November 2025, covering 13,250 hourly bars. Each hour carries end-of-day option chain summaries lagged one full session. An hourly bar therefore never receives that day's closing chain. The target is end-of-day implied volatility averaged across the whole chain's strikes and expiries. It is a single daily number, without a fixed tenor or moneyness.

At each hourly anchor, the inputs include the prior session's chain IV, put-call skew and one-day IV change. They also include a variance-risk-premium proxy, calculated as implied variance minus trailing 21-hour realized variance, alongside price technicals. The model forecasts EOD IV for the session containing each of the next 14 hourly bars.

The premise is familiar and sound. Earnings and macro shocks push NVDA vol up; it then falls over several sessions. Persistence simply carries yesterday's level forward, missing that decay.

The paper pits persistence against a 2-layer encoder-decoder LSTM and HAR-IV. HAR-IV is an OLS regression that predicts forward IV from current IV and its 5- and 22-session means. The test covers 1,963 hourly windows across 281 sessions, running from September 24, 2024 to November 5, 2025.

The paper also has three other results that are straightforward to read. In stratified subsampling of option chains, allocation by lagged cell dispersion brings equal-weighted surface RMSE down to 0.68x SRS at 25% retention. Liquidity-weighted PPS sampling reaches between 0.46x and 0.05x SRS across 25% to 80% retention when evaluated against a liquidity-weighted target. The authors describe that result as holding by construction. Apply the same design to the equal-weighted target and RMSE rises to 2.47x to 3.90x that of SRS.

Short-horizon price prediction goes nowhere. The random walk records $4.148 pooled RMSE, versus $4.156 to $4.161 for the LSTM. An always-up classifier's 55.17% directional hit rate also beats the LSTM's 54.7% to 55.0%. Rolling volume bars fare better on their own task: interior-session excess kurtosis falls to 1.83, compared with 5.00 for time bars and 5.09 for diurnally standardized time bars.

Where HAR-IV earns its advantage

HAR-IV records 1,254 bps RMSE, compared with 1,410 bps for persistence, giving +21.0% MSE skill. Restrict the comparison to strictly future-session targets and skill reaches +22.0%. The pure daily version produces similar results. One session ahead, skill is +17.4% (1,085 vs 1,194 bps); two sessions ahead, it is +22.9% (1,258 vs 1,433 bps).

The top 5 reversal sessions account for 157% of HAR-IV's net squared-error reduction.

Across the remaining 276 sessions, HAR-IV therefore loses to persistence on net. The daily results tell the same story. At 15:00 ET anchors, persistence wins on squared error in 57.3% of sessions (bootstrap CI 50.2% to 64.8%). The median daily loss differential is +2.20, while the mean is −48.8. Against persistence, the session-clustered Diebold-Mariano test returns t = −0.95, p = 0.34. Remove the two largest jump sessions and the result is t = −1.09, p = 0.27.

The authors disclose the concentration and the lack of significance under session-clustered tests, including in their abstract. Their wording in the body is explicit: they "interpret the +17.2% to +21.0% pooled MSE skill as descriptive, episode-concentrated volatility-reversion skill rather than statistically significant uniform daily dominance." That is an honest description. It is also the whole finding for a trader. This forecast behaves like an event model left running every day, paying a small premium through the calm sessions.

Can the LSTM beat three linear terms?

No. The best LSTM reaches 1,283 bps using the parsimonious 20-feature set and a 60% PPS training subsample. That sample is drawn in proportion to an options-activity score, which the paper calls 'Liquidity PPS'. HAR-IV's 1,254 bps is lower. A direct session-clustered DM comparison of LSTM with HAR-IV gives t = +0.51, p = 0.61.

The LSTM depends just as heavily on a few reversals: its Top-5 concentration is 144%. With full-census training, adding eight option-activity aggregates reduces 5-seed mean IV skill from +13.4% to +3.9% (Welch p = 0.011). The paper reports this tier comparison over the same 2024 to 2025 window. We did not find a separate validation step used to choose the 20-feature set. Its standing as the preferred specification therefore depends partly on the test window.

Subsampling makes a case for saving compute, separate from the prediction results. Training with 60% of windows reduces wall-clock time by 21% to 22% and passes TOST equivalence against full census. At 60% retention, PPS preserves 81.2% of top-decile realized-vol windows and Neyman preserves 73.7%, versus 59.9% for SRS. Neither delivers a measurable accuracy improvement. Relative to SRS, IV RMSE differences are +0.51% for PPS (p = 0.608) and +0.09% for Neyman (p = 0.937). The authors make that limitation clear.

A chain average still needs a trade

You cannot buy the full-chain average at its forecast value. Over 2024 to 2025, the index averages 1,820 to 3,972 strike-expiry series, with mean DTE near 246 days. The authors reconstructed it from OPRA close snapshots after an earlier pre-filtered OPRA subset, limited to DTE at most 75 days, introduced a maturity break at January 2024. They repaired that break. Changes in the chain's mix can still move the index level. Post-earnings expiries rolling off provide one example. Composition could account for some of the measured reversion; the paper does not let us determine how much.

Timing adds another qualification. In 21.41% of test pairs, the target is the current session's close and the latest known IV comes from the prior session's close. The remaining 78.59% target the next one or two sessions.

The trading hypothesis is narrow. Following a session in which chain IV rises well above its 22-session mean, a delta-hedged short vega position in a fixed-tenor, near-ATM straddle should capture reversion over the next one to two sessions. Calm sessions should impose a small, measurable cost.

What we are testing

We are building a narrower substitute experiment around an end-of-day HAR-IV extension. We do not test the hourly LSTM or the sampling designs. Nor do we test a trading claim from the authors, who make none. Our options history falls short of their full 2018 to 2025 panel. Without the original vendor feeds or bid/ask quotes, we also cannot reproduce their quote filters or liquidity weights. We instead use a stated EOD IV aggregation over overlapping coverage and explicit transaction-cost assumptions.

The paper's diagnostics determine the priorities. HAR-IV and persistence are the only benchmarks that matter here. We separate P&L on event sessions from P&L on calm sessions, and fix option selection, hedging and spread costs in advance. An edge that disappears when the top five sessions are removed was never a daily strategy. The 157% concentration figure already suggests that outcome. Treat 281 sessions in one stock as a hypothesis, nothing more.