Classification metrics versus trading value
Read accuracy, precision, recall and class imbalance without confusing a correct label with a profitable executable decision.
Identify the record scope, reconcile signals and open exposure, and separate cash flows, research fills and subscriber results.
Research design and arithmetic, not model recommendations. Classification scores are not monetary returns; a registered artifact and a passing software test are not independent strategy validation.
A record can belong to a model's raw signals, a hypothetical strategy, a particular account or a service's copied subscriber route. Begin with that scope, source version and cutoff. A founder's personal account does not automatically describe the model, while a summed family result does not establish one funded account's compounded return. Keep independently verified evidence distinct from supplier-provided summaries.
Ask for original event and accounting records with an inspection route. A revised marketing page may change without preserving its earlier inputs. Record the model identity, instrument scope, period, opportunity definition and performance convention. If those fields are missing, a precise percentage remains a claim with an incomplete denominator. Do not create a league-table position merely because several suppliers publish apparently comparable numbers: they may describe incompatible objects and periods.
Reconcile eligible opportunities to issued signals, abstentions, cancelled instructions, amended calls, unresolved positions and closed outcomes. Explain whether repeated messages attach to one position or create new exposure. A closed-trade win fraction can omit substantial open losses or gains. A model-level abstention rate also matters when accepted predictions alone form the displayed sample.
Request event identifiers and amendment lineage rather than a selected screenshot sequence. A timestamp commitment binds only the bytes covered by the commitment and does not itself establish completeness, fills or subscriber access. Keep deleted or inaccessible records as a known evidence gap. A service that stopped publishing does not disappear from a dated comparison population. Missing outcomes should remain missing; assigning them arbitrary wins or losses would manufacture a different track record.
Imagine an account beginning at 10,000, growing to 11,000 with no external flow, then receiving a 5,000 deposit and remaining at 16,000. The first segment gains ten percent; the second segment after the deposit has zero market return under these assumptions. A flow-segmented time-weighted calculation compounds 1.10 and 1.00 to give ten percent, not the sixty-percent change in raw account value.
The fixture assumes the deposit occurs exactly between the two marked segments. Actual flow timing and valuation require a declared convention and original ledger. Money-weighted and time-weighted returns answer different questions, and summed signal percentages need not equal either. Retain currency conversion, fees and open-position marks. This invented example checks accounting only; it is not a model return, a subscriber outcome or a claim of compliance with any performance standard.
Research may use a stated observation-window convention to select a hypothetical entry reference. Preserve that convention and its directional and timing rules; do not quietly replace it with first-available prices or later favourable prices. Actual broker execution uses execution reports, account identity and quantities. A backtest reference cannot be relabelled as the price every subscriber received.
Inspect quote side, delay, order type, partial fills and costs additional to the prices. A triggered instruction can execute elsewhere, and an unfilled order has no position merely because its price was touched. Keep source model, copied account policy and subscriber route separate. No trading cost should be subtracted twice. When broker reports are absent, the strongest justified assessment may concern research accounting rather than actual execution; state that boundary near the headline rather than hiding it in a footer.
A win fraction requires an outcome definition and eligible count. NIST documents interval methods for binomial proportions, but an independent Bernoulli sampling approximation must not be assumed for overlapping calls without justification. An interval for winning frequency is not an interval for monetary profitability, account return or future drawdown. Payoff sizes and unresolved exposure remain separate evidence requirements.
Ask how many candidate models, windows and exits were inspected before the presented version was chosen. A reserved evaluation region stops being untouched after repeated inspection. Keep revisions visible when entry or exit policy changes. GIPS is linked as a performance-presentation reference, not a certification that a model or this guide complies. This page provides no empirical model panel and therefore invents no uncertainty interval, Sharpe ratio, ranking or probability of trader skill.
Report source completeness, accounting reproducibility, independent evaluation and actual execution separately. Link each finding to a dated artifact and identify the unresolved request. A supplied performance summary can be accurately attributed without becoming independently reproduced evidence. If a correction changes the denominator or cutoff, preserve the earlier version and explain the resulting scope difference.
The blank record export asks for these layers before a reader accepts a verification badge. Reproduce the deposit fixture and inspect a favourable, adverse and unresolved event using the same rules. Do not infer sponsorship, private testing or a paid subscriber experience from a provider note. No consumer-review count or editorial score is fabricated here. A useful assessment states exactly which claim can currently be checked and what additional records would be required for a stronger conclusion.
| Object | Preserve | Do not infer |
|---|---|---|
| Identity | Model, account, source and cutoff | Founder and subscriber records are identical |
| Accounting | Population, flows and open exposure | Raw account growth is strategy return |
| Confidence | Sampling assumptions and selection history | An interval proves future account profit |
Download blank review record (CSV)
Headers only. No trades, subscriber outcomes or operating receipts are supplied.
Documentation establishes the stated mechanism, not tested trading performance. Checked 10 October 2026; a later version can differ.
Read accuracy, precision, recall and class imbalance without confusing a correct label with a profitable executable decision.
Audit availability times, train-only transformations and forward targets before trusting a trading-model split.
Keep candidate selection, independent strategy validation, software checks and cloud production state as separate durable records.
Ask whether a summary concerns probabilities, thresholded label decisions, raw trade percentages or account equity. Each object has a different denominator and source population. A correctly computed Brier loss needs the actual row probabilities; a net equity curve needs positions, cash flows and accounting. None can be reconstructed from a favorable aggregate adjective. The same-table prevalence reference is descriptive, not an untouched benchmark. Keep a revised model's record separate when data, target or transformation changes. An older successful evaluation does not automatically belong to a new marketed version, even when its nickname and founder are unchanged.