Quant Ml Protocol

Quant ML Protocol

This is the general protocol. /model-improve is its Quantamental instantiation; it lives in quantamental/.claude/skills/ and loads only inside that repo. Use it for the mechanics there, but on any conflict of statistical principle (nulls, holdouts, selection honesty), this file wins.

Distilled from two real campaigns (bd-forecast, 9 iterations, Aug 2026; grade-diff, Sep 2026): every rule below was paid for by a specific failure, recorded in references/steering-bank.md. The core insight is a three-tier map of what goes wrong:

• Known unknowns — mechanizable: the checklists and gates below. Run them unprompted.
• Unknown unknowns — partially mechanizable: scheduled adversarial passes against YOUR OWN PROCESS (not just the model). Run these unprompted too.
• Unknown unknown unknowns — NOT self-generatable. Every catch in this tier came from the human: attacks on the decision rule itself, domain knowledge the model lacks, "tied on WHAT?" challenges. The skill's job is to MANUFACTURE the moments where the human can strike: hard gates with the right surface presented and the right questions invited. Never skip a gate because the evidence "seems clear" — clear-seeming evidence is exactly what the gates exist to attack.

Stage 0 — before any code: the protocol file

Instantiate templates/EXPERIMENT-PROTOCOL.md into the project as docs/EXPERIMENT-PROTOCOL.md. Fill in, in this order:

1. Sample-size math FIRST. n_independent = anchors / horizon-overlap. SE(IC) ≈ 1/sqrt(n_indep). Write the detectable-effect floor into the protocol. Cap features at ~n_indep/8. Every later verdict quotes this noise yardstick in the same sentence as its effect size.
2. Instrument mechanics. Tenor/contract holdability measured, not assumed (a front contract that dies mid-horizon cannot carry the trade). Every cost the tradable expression pays — spread, commission, financing × calendar days held, borrow — lives INSIDE the net primary metric from the first backtest. A known cost carried as a caution flag instead of a number has flipped a verdict after promotion (AF2 2026: +3.93 bps/day t 2.74 gross became ~+1.13 at t ~0.8 once the stated 2 bp/calendar-day financing was applied). Execution lag is a cost too: the entry price is the first tradable print AFTER the signal is computable, never the assessment mid the signal was computed from. Measure the lag and apply it from the first backtest; a money number dealt at the mid is a diagnostic.
3. Target decomposition. Subtract everything the market already pays: carry, roll-down, basis. If a naive always-one-side strategy scores well on the raw target, the target is wrong, not the model good.
4. The decision rule, declared. Which metrics are primary (at least two: a ranking metric AND a money metric with costs), what breaks ties (parsimony), and what nulls are: vetoes or caution flags. This rule is itself attackable at the promotion gate. The primary t uses autocorrelation-robust errors (Newey-West or a stationary block bootstrap); a plain-SE t on overlapping or autocorrelated returns is a diagnostic, never the verdict number. Stage 0 is done when every <...> placeholder in docs/EXPERIMENT-PROTOCOL.md sections 0-3 is a real number or an explicit N/A, not before.

Stage 1 — data honesty

• Every feature row carries available_ts; the build FAILS if any feature flunks the point-in-time audit. Publication lag is modeled at JOIN time.
• Raw pulls immutable in dated directories; a re-pull is a new directory.
• Record each source's revision status (vintaged / stored-revised / unknown). For stored-revised sources: start weekly vintage snapshots on day one and stress the top driver with artificial reporting lags NOW, not later.
• Roll-clean any futures/swap series: same-contract changes chained, never spliced levels.

Stage 2 — pre-registration

• Every search is declared BEFORE it runs: axes, cell count, registry hash, and the null design that will judge it. An undeclared run is a diagnostic, never evidence.
• One immutable primary spec per campaign; variants are labeled diagnostics that cannot flip the verdict.
• Every report carries a NOT-DONE table (what was declared but not run, and why). Silence about a skipped check reads as "passed."
• Trial ledger. Keep a running count N of verdict lanes across the campaign — every declared lane that could have produced a pass, plus peeks at accruing lanes used to choose the next build. Every pass is reported next to N × α (expected false passes) and the expected maximum null Sharpe at N trials. A pass inside that noise ceiling is a flag, not a result. N × α is a Bonferroni-shaped bound and over-punishes correlated lanes. When lanes share data, judge the family with a joint bootstrap (Hansen SPA, or Romano-Wolf StepM, which names WHICH lanes survive) and report it beside the N × α line.

Stage 3 — baselines before models

Random walk, carry/hold, always-long, always-short, coin flip — in the TARGET'S OWN SPACE, on the same anchors. Plot them on every cumulative chart. A model that cannot name the naive strategy it beats has no result.

Stage 4 — model discipline

• Simple → complex, one change at a time, champion/challenger.
• Purge: a training row is admitted only after its outcome RESOLVED before the fit date. Remaining overlap gets uniqueness weights.
• Validate the hyperparameter GRID itself (a grid reaching too-strong regularization silently shrinks small specs to the base rate).
• A richer model class earns its place only by BEATING the simple one on honest rows — a tie promotes the simpler model.

Stage 4.1 — the mechanism ladder (every rung tried, or in the NOT-DONE table with a reason and a slot)

Distilled from the grade-diff campaign (Sep 2026), where every rung below was found untried at the gate. Walk it in order; each rung is a declared lane (Stage 2) with its own null bar.

1. The level before the sign. When the target is a level (a diff, a spread, a structure) that shows autocorrelation, fit AR / ARIMA / exponential smoothing / HAR-type models on the level itself. A sign classifier is a try only once a level model holds a verdict row beside it.
2. Magnitude and distribution. Sizing needs a magnitude. Add a regression on the size of the move, a volatility model, and a distributional lane (quantiles or a full predictive distribution) scored with a strictly proper rule: pinball loss or CRPS. Calibration first, sharpness second.
3. Window and recency. Sweep rolling lengths against expanding, and add exponential recency weights as the continuous version; they cost nothing extra. Expanding vs one rolling length is a diagnostic. Small samples usually punish forgetting, so expanding is the prior and a shorter window must beat it on honest rows.
4. Lag structure of outside data. Each outside series enters with declared lags, changes, and interactions with regime state and reset events, within the feature cap. A single z-score per series is the baseline lane, and the ladder starts above it.
5. Feature selection with error control. Decorrelate or cluster first, hold the n_indep/8 cap, then select with model-X knockoffs at a stated false-discovery rate. Knockoffs assume the feature joint distribution is modelled; with autocorrelated features generate them block-wise and label the FDR nominal. Per-feature knockout has no error control and stays a diagnostic.
6. Structure before search. Where the sign of a driver is domain knowledge, encode it as a monotone constraint (LightGBM monotone_constraints, method intermediate or advanced) so the trees cannot fit a noise-driven sign flip. Where one target is under-powered and related targets exist (sibling grades, sibling routes), pool them with a hierarchical model that shrinks each target's coefficients toward the pooled mean by its own standard error (the Vasicek adjustment). Pooling is the largest sample-size multiplier available.
7. Hyperparameters. Optuna/TPE on the winning families, about 50 trials per refit, declared per Stage 4.5 rules. Inner folds of tens to low hundreds of rows mostly fit the folds, so the declared null bar is the judge, not the inner score. Run a zero-search champion beside it (a prior-fitted tabular model such as TabPFN): its k is 1, one ledger lane instead of a search, and a tie promotes it by parsimony.
8. Ensembles across families. Simple averaging and stacking across the families are declared lanes. Prospectively, exponential-weights aggregation over the live specs replaces champion/challenger selection: the weights adapt causally, and the selection haircut becomes a regret bound of order sqrt(T log N) against the best single spec.
9. Late-lane data. A series that starts after the sample start is tested in EVERY window configuration the main lanes use, not one, and its thin-window status (Stage 5.5) is stated beside each result.
10. Execution lag and real costs. Stage 0 item 2, applied: the rung most often "already flagged" at the gate with the metric still dealt at the mid.

Stage 4.5 — the coverage interrogation (fire at yourself before every gate and every report)

Keith's questions (2026-09-05, prc26), asked verbatim and answered in writing, one table row each, every answer carrying a pointer (protocol row, ledger row, report file). "Not tried" is a valid answer only with the reason and a scheduled slot.

1. What have you tried?
2. What features?
3. What feature engineering?
4. What model?
5. Feature and model parameters tuning?
6. Optuna optimisation?
7. Are the agent being detailed enough?
8. Full effort used to find suitable model?
9. What is missing that we should try?

Rules for the answers:

• A "tried" carries its verdict row (id, numbers, paired interval) or it is not tried. A run with no declaration is a diagnostic, not a try.
• Answer 1 walks the Stage 4.1 ladder rung by rung; an unwalked rung is a NOT-DONE row, never silence.
• Answers 5 and 6 name the search: cells, k, the optimism bound, the CV design. "No search" is a finding to report, never a silent default. Optuna or any sequential optimiser is a declared search like any other: budget, seed, objective and the null bar it is judged against are declared BEFORE the first trial, and its k is the trial count, not 1.
• Answer 9 lists at least one untried INPUT (new data), one untried MECHANISM (model class, target treatment, loss) and one untried TUNING, each with a cost estimate and a draft declaration. The user picks; the agent does not run them unasked.
• Answers 7 and 8 are graded by the user, not the author. The author's job is to make grading possible: this table, the NOT-DONE table and the ledger line, side by side. Self-grading them "yes" is banned.
• The table rides every campaign report and the Stage 6 surface (item 6).

Stage 5 — the validation gauntlet (adversarial, process-level)

1. Walk-forward only. No random splits on time series, ever.
2. Process nulls, not model nulls: re-run the ENTIRE recipe — search, selection, tuning included — on time-shifted (information-free) inputs, ≥20 replays. Compare the real result to the null DISTRIBUTION of the process. A wide search manufactures large effects from pure noise; the null bar for a search is what the search finds in nothing. Fewer kept replays than the declared floor = no null = NO verdict; thin the null, lose the lane. Count the null's EFFECTIVE draws: placements or replays that share most of their data (e.g. overlapping windows on a circular clock) are one draw, not N — a p95 over near-duplicates is not a 5% tail. Prefer a stationary block bootstrap of the return series when placements must overlap.
3. Selection-honest hold-out: choose the spec on early years only, judge on untouched late years. Report the haircut (as-run minus honest) as its own number.
4. Leave-one-year-out; per-feature knockout; seasonality controls (calendar dummies alone as a baseline); revision stress on every stored-revised driver. A knockout answers "does the model need X", not "does X move the target". For the second question (a turnaround, a chokepoint count, a programme change) use double/debiased ML: orthogonalise driver and target against the nuisance features with cross-fitting, and report the effect with its interval.
5. Thin-window rule: any candidate evaluated on n < 300 anchors goes to the pre-registered ACCRUAL LIST with a scheduled re-test date — never into the model, never forgotten. The floor counts the anchors the VERDICT statistic rides (judged rows after any tune/judge split), not the full sample.
6. Mandated self-audit before any verdict: answer "what else is wrong with what you did?" with a numbered list of findings covering ALL FOUR categories — data, statistics, code, process. A category with no finding needs a written one-line defense of why it is clean; there is no minimum count (coverage, not quota — a quota manufactures filler findings). One pass, written down, in the report.
7. Independent process critique before any Stage 6 surface: a reviewer that did not build or run the campaign (different agent or model, given artifacts not summaries) attacks the method — nulls, multiplicity, cost model, selection. Its findings ride the gate surface verbatim. The author never reviews alone; the same hands authored, ran, and judged everything else.

Stage 6 — THE PROMOTION GATE (hard user gate — never skip)

Promotion, spec changes, and "best model" verdicts STOP here. Present, in one surface:

1. The honest table (nominate-early / judge-late) AND the as-run numbers, side by side, gap stated.
2. EVERY metric with a paired interval — never a single-metric verdict. ("All tied" on one metric has been overturned by the money metric.) Include the trial-ledger line: N trials this campaign, expected max null Sharpe at N, observed Sharpe beside it, and the SPA or StepM result when the lanes share data (Stage 2).
3. The decision rule restated, with the sentence: "This verdict is only as good as this rule — attack the rule, not just the numbers."
4. All caution flags, openly (a null that fails is demoted to a flag only if the declared decision rule says so — and it rides every future report).
5. Then ask the user the steering bank's gate questions (references/steering-bank.md §Gate) and WAIT.
6. The coverage interrogation table (Stage 4.5), so the user can grade the detail and the effort, not just the verdict.

Promotion requires explicit user sign-off, every time. Being unattended never authorises skipping the gate — ship the surface as the deliverable instead.

Stage 7 — prospective discipline (the real judge)

• Append-only ledger: one row per scheduled run, entry recorded at run time, judged on its own date, NEVER edited. Old-spec rows are archived and labeled, not deleted (deleting one row makes every remaining row worthless).
• Spec tags on every call; a spec change mid-stream is two track records.
• Numeric tripwires at declaration. Every prospective lane declares, before its first row: a REVIEW threshold and a KILL threshold derived from the backtest's own return distribution, plus an operational kill (skip/error rate). "Judgement stays with the user" gates promotion; it is not a substitute for tripwires. A lane too thin to CONFIRM its edge at its accrual rate is declared kill-only, with the detectable- effect horizon stated. The ledger is peeked at every run, so fixed-n tests on it are invalid. Judge tripwires with anytime-valid tools: an e-process for the kill test and a confidence sequence for the edge estimate. Both stay valid at every stopping time, so daily looks keep the declared error rate.
• Sizing shrinks for estimation error. The edge that sets the size is itself an estimate from a thin ledger. Shrink the size scalar (Kelly fraction, vol-target multiplier) by the standard error of the edge (Baker and McHale 2013 give the back-of-envelope form for the win probability) and state the shrunk size beside the raw one.
• Operational reality is part of the model: what time it runs, what price the entry is, what the backtest's marks actually are (settle vs intraday), what happens on holidays — measured, not assumed.
• Every surface a consumer sees must have every element explainable ("why is this row here?" always has an answer), with honest-limitation boxes printed, not hidden.
• Scheduled re-tests for the accrual list; drift monitoring vs the backtest's own distribution.

Cadence of self-audits

Run the Stage 5.6 self-audit: before every verdict, and after every campaign.

References

• references/steering-bank.md — the human steering-question bank (fire them at yourself per stage; fire the Gate set at the user), plus the genericized war stories behind every rule above.
• templates/EXPERIMENT-PROTOCOL.md — the protocol file to instantiate at Stage 0.
• references/physical-diff-addendum.md — READ when the target is a physical differential to a PRA benchmark (grade diffs, OSP diffs, regional spreads): regime register, reset-clean targets, min-operator benchmarks, structure decomposition, physical expression cost, domain baselines, non-statistical tripwires.