A 250-tree random forest built on 21 engineered features is the strongest part of this paper, adding +0.158 mean per-day correlation. The 200M-parameter backbone remains in place across all 12 variants, so its contribution is never isolated.
Dewage, De Silva and Mondal freeze TimesFM, a decoder-only transformer with 200M parameters, and use it to forecast the next 60 one-minute returns of a US stock between 9:30 and 10:30 AM. Its context covers the previous regular session and the current morning's premarket from 4:30 to 9:29. Those inputs become 300 one-minute bars across 7 channels: normalized OHLC, log volume, bar-to-bar momentum and intrabar volatility.
TimesFM alone contributes little. Mean per-day correlation is 0.0586, pooled correlation is 0.0614, and cross-day correlation is minus 0.0357. Its daily-direction result also trails a historical-mean baseline fixed at exactly zero.
The authors leave the backbone frozen and add corrections around it. AttnCorrect, the first neural corrector, applies two transformer layers with four heads to the premarket sequence and has about 471K trainable parameters. GatedLinear takes a different route. It compresses the 300x7 premarket tensor through a low-rank bilinear projection into 32 numbers, then concatenates them with 21 hand-built multiscale summary features and TimesFM's own 60-step forecast. A gated residual comes out, using about 49K parameters. Finally, a Random Forest with 250 trees, depth 12 and minimum leaf 3 learns from the residuals of that combination using the 21 summary features. Frozen backbone, neural delta and forest delta are added to form the forecast.
The data consists of 1-minute bars for ten mega-cap technology names: NVDA, MSFT, AAPL, GOOG, GOOGL, AMZN, META, AVGO, TSLA and NFLX. It runs from December 2024 to January 2026 and contains 2,011,399 rows. Each stock receives a chronological 186/40/40 day split, while NFLX uses 104/22/23 because its history is shorter. Evaluation therefore covers 40 test days for each stock and 23 for NFLX.
GatedLinear plus Random Forest posts mean per-day correlation of 0.3730 and pooled correlation of 0.5972, compared with 0.0586 and 0.0614 for the frozen backbone. Return RMSE is 0.1535%, with MAE of 0.1079%. The 0.3730 figure averages results across the ten names. The paper ablates twelve variants. No strategy is traded: the reported outputs are correlation and error metrics, and the authors say transaction costs and market impact are not modeled.
The forest carries the result
Remove the forest and performance gives way. GatedLinear drops from 0.3730 to 0.2147 per-day and from 0.5631 to 0.3058 cross-day. Its RMSE rises from 0.1535% to 0.1710%. AttnCorrect declines from 0.3678 to 0.2335 per-day and from 0.5819 to 0.3829 cross-day, while RMSE moves from 0.1547% to 0.1740%. Using the paper's direction of comparison, adding the forest reduces RMSE by 10.2% and 11.1% respectively.
The authors present the two halves as close contributors. The abstract describes classical residual learning as "matching or exceeding the neural correction component"; the conclusion states "Classical and neural components contribute nearly equally"; and the first practical recommendation says the forest "provides contributions nearly matching". Table V gives a more uneven picture. For GatedLinear, the forest adds +0.158 per-day and +0.257 cross-day. Self-attention, the strongest neural component in the table, contributes +0.0988 per-day and minus 0.035 cross-day. The neural corrector as a whole remains untested in isolation because none of the 12 variants combines frozen TimesFM with the forest while omitting a neural module.
They looked for the answer and published the result they found.
The plain LSTM baseline reaches 0.3519 per-day and 0.4943 pooled. Both hybrids without the forest trail it: GatedLinear-NoRF records 0.2147/0.3219 and AttnCorrect-NoRF 0.2335/0.3698. A plain MLP baseline also beats both, with 0.2407 per-day and 0.4584 pooled. Before the forest arrives, a frozen 200M-parameter foundation model paired with a trained neural corrector loses to an ordinary LSTM on both correlation measures. The LSTM's RMSE, 0.1744%, is only marginally worse than their 0.1710% and 0.1740%.
The paper also reports that two of the three pieces in its "best" architecture cause harm when the forest is absent. Removing the bilinear projection raises per-day correlation by +0.128, from 0.2147 to 0.3422. RMSE improves 6.4%, moving from 0.1710% to 0.1600%. Removing the gate lifts per-day correlation by +0.072 to 0.2864. The authors attribute this to the 65x compression from 2,100 dimensions into 32, which discards recent premarket dynamics. The forest reads the summary statistics directly and compensates. Their account is plausible, and it leaves the winning configuration dependent on a tree ensemble over 21 engineered features to repair damage introduced by the preceding neural block.
A decisive control is missing from the 12-variant table. The paper includes neither a Random Forest alone on the 21 features nor an LSTM plus RF. Without one of those tests, the results cannot isolate what the frozen 200M backbone adds beyond a forest trained on premarket summaries. The Limitations section acknowledges this in one line: the work "evaluates the hybrid correction methodology rather than claiming that TimesFM itself is suited for financial prediction tasks". That framing is candid. Yet the title, abstract and practical recommendations emphasize foundation model adaptation, while the surviving ablation evidence favors engineered features and trees.
Which correlation reaches the trade?
The abstract places 0.597 pooled correlation beside the 6.4x per-day improvement in the same sentence. Later, the authors explain that pooled correlation flattens every prediction and observation into a single vector and may be inflated by cross-day variance structure. A model can identify more volatile days and produce nonzero pooled correlation without any within-day skill. They describe mean per-day correlation as more conservative and more relevant to intraday trading, then recommend reporting all three metrics in future work. Their own diagnosis bears directly on their headline figure.
By that standard, 0.3730 within-session is the figure to retain rather than 0.597. Results vary sharply across stocks. NVDA moves from 0.018 to 0.473 per-day while reaching 0.739 pooled. AAPL manages only 0.171 per-day against 0.347 pooled. The authors interpret that gap as stronger capture of cross-day variance than within-day pattern among quieter names.
The test design limits what these figures can establish. Every result comes from one contiguous window at the sample's end: 40 test days for each stock and 23 for NFLX, without a walk-forward exercise or a second regime. All 12 variants are ranked on the same window, making "best overall" a test-set selection. Every run uses seed 42, and the paper reports no standard errors or t-statistics. Per-day component differences from 0.072 to 0.158 and cross-day differences from 0.035 to 0.257 therefore come without sampling error. The authors further acknowledge that the 6.4x comparison begins from a baseline of 0.059, noting that limited baseline correlation lets even modest absolute gains produce large relative improvements. The universe contains ten correlated mega-caps, including the near-duplicate share classes GOOG and GOOGL, so its effective cross-section is well under ten names.
Our directional test
We built the directional version absent from the paper. We compounded the 60 one-minute forecasts, took a long position when the compounded return was positive and a short position when it was negative, entered at the observed 9:30 open, exited at the observed 10:30 open and equal weighted the positions. This trading rule is ours.
The following figures come from our own run. Over 2024-01-01 to 2026-07-31, using nine names because GOOG did not resolve, total return was minus 0.55% and Sharpe was minus 0.42. The run contained 570 round trips, produced a 45.16% win rate and had a profit factor of 0.93. We charged $0.004 a share with a $1 order minimum. Annualized volatility came to 0.51%, and maximum drawdown was minus 0.80%. The book carried little risk, while a 0.93 profit factor indicates a near-break-even gross edge consumed by commission.
The comparison requires care because the measurements differ. The paper reports mean per-day correlation of 0.3730 and pooled correlation of 0.5972 before costs. Our result is net P&L: minus 0.55% total return and minus 0.42 Sharpe over 570 round trips after commission. These are different quantities, and neither tests the other.
Our implementation is the leading source of doubt. Training used MSE alone, omitting the directional (0.3) and cumulative (0.2) loss terms specified in the paper. Those terms target sign and cumulative return, precisely the properties monetized by our trade. The paper fully pins down only 13 of the 21 multiscale coordinates, so we defined the remaining 8 ourselves. Those features are the forest's sole input, and the ablation assigns that forest +0.158 per-day and +0.257 cross-day. Our window and 70/15/15 split also differ from the paper's 186/40/40. This difference works against using it as an excuse because trading across dates seen by the model should favor our outcome. We cannot reconcile the gap from the information available, and we do not claim that the authors' correlations are wrong.
The metric-to-P&L problem does carry across. The authors say a pooled correlation of 0.597 can arise from volatility structure, and such a result is compatible with a coin-flip directional call. Our 45.16% win rate looks roughly like one. We have raised the same objection before (/articles/where-finsmart-s-returns-come-from). Correlation pays only after it becomes a position size and a fill.
One additional table could change my view using code the authors already have: Random Forest on the 21 features alone, with no TimesFM and no neural corrector, evaluated on the same test days. Our threshold is 0.10 of per-day correlation. If the frozen backbone adds that much beyond the forest, the foundation-model framing earns its title.