When Does Model Complexity Pay Off?

Higher AUC did not reliably translate into better decisions in this exploratory comparison. In a bake-off of six model families on one noisy, short-horizon prediction task, the five per-stock sequence, board-transformer and graph-attention configurations out-ranked the three gradient-boosted configurations on the headline label, by 0.016 to 0.030 over the LightGBM baseline. The higher AUC was accompanied by higher top-1% precision in several configurations, but the observed differences were smaller than the variation seen across training seeds in a related evaluation; the sequence models' top picks carried one and a half to three times the adverse-event exposure of the trees'; and the one context feature that seemed to fix the tail turned out to be a leak. Rigorous evaluation mattered more than another architecture. This is an exploratory study, not a benchmark; the controlled version follows at the end.

Key findings

• Five per-stock sequence, board-transformer and graph-attention configurations exceeded the three gradient-boosted configurations on headline AUC, 0.595 to 0.609 against 0.573 to 0.583, with one training seed each. The two foundation-model configurations and the text model, at 0.584 to 0.593, also exceeded the best tree, by 0.001 to 0.010.
• Among each model's top 1% of candidates, precision ran from 20% to 31%, higher in several sequence configurations than for the trees, but the differences were smaller than the 14-point spread across three training seeds of one model.
• Higher AUC coincided with one and a half to three times the adverse-event exposure among the top picks, measured by the crash label: 36% to 61% for the five sequence and board configurations and 65% for the text model, against 18% to 24% for the trees, on a population rate of 7%.
• The largest apparent improvement in the study, from an issuer-country feature, was a leak from a reference table that was not point-in-time, and was retracted.
• Related experiments with multiple seeds and null models exposed the limitations of single-run comparisons, while the main bake-off used one seed per configuration. Reading the table needs that context: without it, any family could be shown to win.

1. The question

"Which architecture is best" changes its answer with the training seed. The question an ML engineering leader has to answer is different: does a sequence model earn its cost on this task? The cost is more than parameters. It is a slower training loop, a serving path that must carry state, a model that is harder to audit, and a result that is harder to reproduce. Over a year of work on one-second market data, architecture proposals arrived faster than any other kind, and by October 2026 twelve families had been tried. This article reports the one comparison that ran a set of them on one task with one judge, and what the evaluation had to include before its answer could be read at all.

2. Experimental setup

Data and population

The data is the premarket session of US equities at one-second resolution with the national best bid and offer. A decision is scored every second for every stock that is visible in the top five by volume, has a fresh, sane quote with a spread of at most 1%, and is trading at least 100,000 shares a minute, between 04:01 and 04:25 Eastern. These are the candidate seconds: 867,894 of them over 685 days, of which 209,576 fall in the judged year.

Targets

Each label is read from a fill at the ask one second after the decision, on the path of the bid, so the spread is inside the label:

• +5% before −3% within 300 seconds (base rate in 2026: 23.6%)
• +8% before −3% within 300 seconds (14.9%); the headline label
• +10% before −4% within 300 seconds (12.9%)
• crash: the bid falls at least 8% inside some 10-second span within the next 120 seconds (7.2%)

The first three resolve at whichever side is touched first. The crash label is independent of them, and a second can carry both a positive headline label and a crash; section 3 returns to this with timestamps.

Inputs

57 per-second features of momentum, smoothness, price levels, order flow, freshness and the state of the board, served from a feature store that certifies every feature group by truncation before it is written: each feature recomputed on a prefix of the day must equal its value computed on the whole day. That certification is a truncation test: it proves that no feature depends on observations after the decision second. Whether an external reference value was knowable on the day is a different property, historical availability, and section 4 shows what happens when it is not checked. The tabular models read the features at the decision second. The sequence models read the same features' history from 04:00. The foundation models read the mid price on a 5-second grid. The text models read a typed rendering of the same state.

The unit of a sample, and what "same judge" means

The unit is the candidate second. A sequence model sees one stock-day's window from 04:00:00 to 04:24:59, 1,500 seconds, as one sequence and emits a score at every second; 2,385 training sequences are 2,385 stock-days. The tabular models train on the 165,254 candidate seconds inside those same stock-days. In the judged year every configuration in the main comparison scores every one of the 209,576 candidate seconds of 2026 and is judged on the same seconds, joined by date, symbol and second, against the same labels; a check of the score files confirmed full coverage for each of the eleven rows in the main table. One further contender, a 421M-parameter text checkpoint, was scored on every tenth second for cost and is reported separately in section 3, outside the main table.

"Same judge" means the same seconds, the same labels and the same metric code. It does not mean a controlled comparison. Most contenders ran with one seed. There was no fixed compute budget. The networks were early-stopped on the validation months; LightGBM ran a fixed 500 trees at a learning rate of 0.03, 31 leaves, at least 300 rows per leaf and 0.8 subsampling, untuned. The sequence networks are small: width 48, three layers, eight epochs at a learning rate of 0.002, seed 0.

Split and causality

The split is chronological and the newest year is judged, since the market of 2026 is the one a model would serve: training 2024-01 to 2025-06, early stopping on 2025-07 to 2025-08, 2025-09 to 2025-12 kept as the history the day thresholds need, and 2026-01 to 2026-09 judged. Every trained network has a causality test: its output at second t must be identical when the day is cut at t , with the stock's own history and the rest of the board after t removed. The cross-stock features and edges are tested the same way.

The contenders

3. Main results

The 2026 judge over the same 209,576 candidate seconds for every row. Precision and crash share are computed over each model's top 1% of candidate seconds by its headline score.

Float caveat: the three rows marked + float, including the highest-AUC row, read a float that is missing from the day files for the judged days from 2026-08-12, so they carry a time-aligned missing-data pattern, and the float's historical availability rests on the provider's as-of records (section 4). Read those three rows with caution; the other eight are not affected.

Ranking

The five per-stock sequence, board-transformer and graph-attention configurations out-rank every tree configuration on the headline label: 0.595 to 0.609 against 0.573 to 0.583, which is 0.016 to 0.030 over the LightGBM baseline and 0.012 to 0.026 over the best tree. The two foundation-model configurations (0.591 and 0.584) and the text model (0.593) are also above the best tree, by 0.001 to 0.010. The ordering of the five held on a narrower population the study was repeated on, where the gap was wider (0.58 to 0.59 against 0.54). The observed ranking advantage was consistent across the five sequence and board configurations and two evaluation populations, but its statistical reliability remains unestablished. It is also small, and it is not what the decision uses.

The top of the ranking

Precision among the top 1% of seconds runs from 20% to 31% across the main table: the trees 21% to 28%, the sequence and board models 26% to 31%. Several sequence configurations are higher than the baseline's 22.7%, and the best, 31.3% for the board transformer with float, is above every tree. What is not established is that the differences are reliable: the spread between training seeds of one model, measured in section 4, is 14 points, so 31.3% is not distinguishable from the trees' 28.1% on this evidence.

Why higher AUC did not mean safer decisions

Adverse-event exposure is the fraction of a model's top-1% seconds whose crash label is true: the bid fell at least 8% inside some ten-second span within two minutes of the fill. It is an exposure, not a realised loss; the example below shows a pick that reaches its target and then meets the crash. The trees' top picks carry it 18% to 24% of the time; the five sequence and board configurations' top picks 36% to 61%, and the text model's 65%, against a population rate of 7%.

The two labels overlap because they are read from the same fill. A schematic example with the timestamps a reader can check against the definitions: the decision is made at 04:05:00 and the fill is the ask one second later, $4.00 at 04:05:01. The bid reaches $4.32, up 8.0%, at 04:05:41, forty seconds after the fill and inside the 300-second window, so the +8/−3 label is true. Between 04:06:01 and 04:06:11 the bid falls from $4.40 to $4.03, a drop of 8.4% inside one ten-second span that ends 70 seconds after the fill, inside the 120-second window, so the crash label is also true. For the +8/−3 bracket the pick is a win, since the target was touched first; for a position still open at 04:06:01, or one entered a few seconds after the first, the collapse is what it meets. A top pick can therefore count toward a 31% precision and a 61% adverse-event exposure at once. The exposure is per second and concentrated in a few stock-days, since the top-1% seconds of a day cluster together.

The sequence models rank by momentum in the tape, and at this horizon the fastest rises are the ones that collapse. The trees' top picks are flatter: lower AUC, less exposure. The bracket books built from each model's scores were not better for the sequence models than for the trees despite the AUC, and this tail is the reason. AUC on the entry label is the wrong objective on its own for a decision that holds a position afterwards.

Scored on a tenth of the seconds: the 421M checkpoint

A 421M-parameter open decision-model checkpoint, a ModernBERT encoder with decision heads, was fine-tuned the same way as ModernBERT-base (top eight layers and the head) and scored on every tenth candidate second for cost: 21,052 of the 209,576, with the same label rate (14.7%). Its sampling differs from the table's, so its metrics are not comparable with it and are given on their own: AUC 0.552 on the headline label and 0.730 on the crash label, top-1% precision 18.0%, adverse-event exposure 20.9%. On its own population it was the weakest contender. The finding that a text rendering of the state scores below the numeric models stands on ModernBERT-base, which scored every second.

4. Why the results are difficult to trust

Seeds

On the narrower population, three training seeds of the same RWKV gave headline AUC of 0.589, 0.573 and 0.579 and top-1% precision of 21.8%, 32.7% and 36.0%; three seeds of the same GRU gave 0.582, 0.566 and 0.575 and 29.3%, 26.0% and 27.6%. AUC moves by 0.016 with the seed; precision at the top moves by 14 points, more than any difference between families in the main table. A single run of any family can be made to beat any other family by choosing the seed. Every row in the main table is a single run.

Periods

Variation across evaluation periods is not measured here: there is no confidence interval over time blocks or stock-day clusters, and seed variation is not a substitute for it. One number says the period matters more than the seed: the sequence models reached 0.66 to 0.70 AUC on the early-stopping months of 2025 and 0.58 to 0.61 on 2026. The market moved between the two. A model selected on the validation number would have been judged by a different market than the one it was trained for.

Correlated samples

On a related dataset from the same tape, 540,000 candidate rows came from 1,189 stock-days, and consecutive rows' outcomes correlated at 0.89. A tree leaf of 300 rows can be one stretch of one stock-day, so any feature that varies smoothly within a day lets a model memorise training stretches. Adding realistic features made an out-of-sample ranking worse; the same features shuffled between stock-days of the same month hurt almost as much; pure noise was ignored. The effective number of independent observations is far smaller than the number of candidate seconds, and stock-days are the appropriate starting unit for blocked uncertainty estimates, with the caveat that stock-days are correlated too. A decision-level score over 60 to 140 picks swung under null models by more than any architecture difference above. That is why the study that follows judges by population ranking with a shuffled-between-units null and a noise null before any decision-level number is quoted.

The result that looked like a win, and was a leak

A second round added two static channels per stock: the float and the issuer's country. With them the adverse-event exposure of the top picks fell sharply, RWKV from 36% to 20% and the board transformer from 48% to 29%, with no change in hit rate. It read as the sequence models learning which names collapse. It was retracted the next day. The country came from a current company profile, not a point-in-time one, so a stock whose profile no longer existed was marked unknown, and unknown predicted the outcome: survivorship, not signal. With the country channel removed the crash cut vanished entirely (RWKV 36% to 46%, the board transformer 48% to 61%; those are the figures in the table). A context feature that moves the tail but not the hit rate is the first thing to audit, and the reference table it comes from has to be point-in-time or the whole arm is void.

The float arm that remains has its own caveat. For days through 2026-08-11 the float is the provider's shares outstanding as of the previous trading day, the vendor's historical record rather than a capture made at the time, chosen so that a filing during the session cannot leak in. For days from 2026-08-12 the day files held no float at all, so the float channel reads unknown for the last seven weeks of the judged year, and the float arms carry a missing-data pattern aligned with time. A point-in-time rebuild of those days exists and had not been applied to the shared files when the bake-off ran. The float arms are reported with that caveat; it does not touch the other eight rows.

The decision threshold

Two thresholds appear in this article, and they are different objects. The idealised break-even precision of the +8/−3 bracket, with the spread already inside the labels and no other cost, solves 8 p = 3(1 − p ), so p = 3/11 = 27.3%. The 32% used elsewhere in the program is a scenario-specific, empirically estimated threshold from simulated bracket books: it includes the overshoot of a stop that fills at the next second's bid (the books' median loss was −3.1% to −3.4%) and the picks that touched neither side and were sold at the bid at 300 seconds, under the book's rules of one entry per stock per day and at most three open positions, and it excludes fees. The top-1% precision statistic is computed on candidate seconds and applies none of those rules. Read against it, two configurations exceed the idealised 27% (31.3% and 28.1%), by less than the seed spread, and none reaches the book-estimated 32%; neither statement establishes a book outcome in either direction. Only a simulated book with the fills, the position limits and the overlapping events decides actual decision performance, and the bracket books mentioned in section 3 are that object, not the precision column.

5. What survived scrutiny

• In the main comparison, on this task and these features, the five sequence and board configurations rank better than the three tree configurations by 0.016 to 0.030 AUC over the baseline, consistently across the five and on two populations, with one seed each. Top-1% precision was higher in several of them, by less than the seed spread.
• Higher AUC co-occurred with higher adverse-event exposure among the top picks in every sequence and board configuration tested; the exposure is the crash label's rate among the picks, not a realised loss.
• Attention and edges across stocks added nothing measurable over the per-stock model on these features at one second, in this study and in one later ablation (appendix A).
• Zero-shot foundation-model forecasts of the next five minutes added nothing over the features in this comparison.
• A text rendering of the state scored below the numeric models that read the same information, on one fine-tuned encoder scored on every second.
• A static context feature from a non-point-in-time table produced the largest apparent improvement in the study, and it was a leak.
• Preliminary, and not tested as its own hypothesis: a sequence encoder's hidden state, used as the state of a learned exit rule, was the only configuration whose out-of-sample behaviour held across the months judged. It is seed-fragile.

6. Engineering implications

Extra model complexity may be justified when the answers to these questions are yes, and each can be checked before training the expensive thing:

1. Does the tabular summary lose information? Train the sequence model on the raw history of the same features and the trees on the summaries. Here the gap was 0.016 to 0.030 AUC over the baseline, real but inside what a seed does to the top of the ranking.
2. Is there room between the base rate and the precision the decision needs? Here the base rate was 15%, the idealised break-even 27% and the book-estimated one 32%; the best precision reached was 31% with a seed spread of 14 points. No architecture closes a gap the information does not allow.
3. Does the top of the ranking carry the tail you will hold? Measure the adverse label among the top picks, not only precision. Here it doubled or tripled with the sequence models.
4. Does cross-entity structure carry signal at the decision cadence? Ablate the edges before believing them.
5. Is there a control problem downstream? An exit, a stop, a sizing rule. The preliminary evidence here is that a sequence model contributed as a state encoder for that problem, not as a scorer.
6. Can the comparison be read at all? Several seeds, a population-level metric, two null models, uncertainty over time blocks, and the newest period judged. Without these, any family can be shown to win.

What it cost

Inference latency was not measured, and on a per-second serving path it is the cost that decides deployment. The controlled study measures it.

7. The controlled study that follows

The bake-off above is exploratory: one seed per row, no fixed compute budget, no latency, architectures run as they were proposed. The study that follows is the controlled version, built to be reproducible outside the platform, and it separates the two claims this article leaves standing into two hypotheses tested on their own:

• H1, as scorers: sequence models' AUC gain over trees on point-in-time tabular features survives five seeds and blocked confidence intervals, and does or does not translate into top-k precision, calibration and adverse-label exposure.
• H2, as state encoders: a downstream decision rule over a sequence encoder's state out-performs the same rule over the tabular state, out of sample, across seeds and periods.

Deliverables: a research report, a public benchmark repository, and a readable summary of what was found.

Appendix A: secondary experiments

Four later experiments on the same tape point the same way and are listed here rather than argued from.

• Four backbones on a longer window. On a 600-second oracle label with a board graph, Griffin, Mamba-2, RWKV and xLSTM gave 2026 AUC of 0.781, 0.782, 0.777 and 0.778. Cutting the board edges changed AUC from 0.787 to 0.786 and the top-1% hit rate by under a point.
• A neural controlled differential equation over the last 120 seconds of tape, on another target: AUC 0.62 to 0.67 against LightGBM's 0.63 to 0.68, and its top picks hit less often than the trees'.
• Retrieval. A nearest-neighbour case memory over 156,000 past moments, with hand-made and learned similarity, scored at or below LightGBM on the same rows.
• AutoML. Twelve families with tuning and soft blends under time-ordered cross-validation found no cell that was positive in all three walk-forward windows.

Appendix B: experiment manifest

Scope

This is an exploratory comparison on one platform's data, and it is complete on its own terms: one task, one judge, every contender under a causality test, every number transcribed from the run's log and traceable through appendix B. What this exploratory comparison documents is the observed ordering of the tested configurations in these runs, the adverse-event exposure hidden by headline AUC, the leakage uncovered during evaluation, and the limitations that a controlled study must address. What it leaves to the controlled study are the fair tuning, the seeds and the blocked confidence intervals that would turn the ordering from an observed effect into a measured one, and a public dataset on which anyone can repeat it.

The experiments were implemented by AI coding agents working inside the platform's causality gates; the research questions, the audits and the retractions were mine. The leak above was built by that process and caught by it, which is the argument for the gates. The numbers are model-quality numbers, not trading results. The findings are about these features, these targets, these architectures at these sizes and these periods; the checklist in section 6 is what I would test next.