Documentation

The site tests trading strategies against a curated catalog of 63 stress cases — 31 historical market episodes (Lehman 2008, COVID 2020, Luna 2022, etc.) plus 32 synthetic stress probes (50 Monte Carlo replicas each) — and reports per-regime performance via a 10-axis failure-mode taxonomy. As of May 2026 V_RECOVERY is treated as a diagnostic path pattern and decomposed into SHARP_CRASH and TREND_UP for score aggregation, so the score uses 9 effective buckets.

The fastest way to get oriented:

1. Open a strategy report. Browse the pre-computed reports listed on the home page, or open one directly via its URL — for example /s/sma-200-crossover or /s/rsi-30-70.
2. Read the headline numbers. The robustness score (0–100) and Worst-Case Drawdown at the 95th percentile (WCDD95) summarize the strategy across all evaluated cases.
3. Inspect the failure-mode breakdown. The Scenario Dashboard groups the 10 failure modes into 6 buckets (Rising, Sideways, Calm, Volatile, Falling, Crash). Selecting a card pins a representative case in the Inspector below.
4. Review the per-scenario activity. The Per-Scenario Activity table lists every case the strategy was tested against — empirical anchors plus synthetic probes — including the number of trades, the realized return, and the maximum drawdown.

Every strategy report follows the same layout. From top to bottom:

Many trading indicators depend on prior price history. A 200-day simple moving average, for example, is undefined for the first 199 bars and only stabilizes after sufficient further data. In a stress case of 126–252 trading days, an indicator launched from an empty state would produce no actionable signal until most of the case has already elapsed.

Synthetic stress cases on this platform therefore begin with a 250-day antecedent pre-period followed by a short transition window before the stress regime starts. The pre-period exists so that long-lookback indicators have stable values when the stress phase begins. Performance attribution — trades, returns, drawdowns — is computed only from the stress phase onward; the pre-period is used solely to initialize indicator state.

The antecedent regime itself is not the stress regime. It is drawn from an asset-class-typical distribution (e.g. for equity assets: predominantly bull-trend or sideways-low-vol, with a smaller share of weak-bear) so that indicators enter the stress phase from a representative, rather than artificial, prior state.

Antecedent regimes are deterministically assigned per simulation seed, ensuring reproducibility. Results may depend on the assumed antecedent distribution, which defines the initial market state prior to stress onset; this is a modeling choice and should be considered when interpreting results. Cases where the warmup requirement still exceeds the case duration (e.g. very long lookback windows on short cases) are excluded transparently and shown in the report header.

Empirical historical cases (Lehman 2008, COVID 2020, etc.) carry their natural pre-period from the surrounding real market data and do not require this construction.

The robustness score (0–100) aggregates the strategy’s behavior across all evaluated cases, weighted by failure-mode coverage. It is a relative measure of consistency — higher reflects more uniform behavior across regimes; lower reflects substantial regime-dependent variability.

The score is not a risk measure, a return forecast, or a recommendation. It cannot be compared in absolute terms across different case catalogs.

The aggregated score should be read alongside the per-failure-mode breakdown — two strategies with the same score can have very different regime profiles.

WCDD95 is the 95th-percentile worst-case drawdown — in 95% of observed scenarios the strategy’s drawdown did not exceed this level. It represents downside exposure under the tested stress conditions.

Does not imply future performance. The metric summarizes what was observed in the case catalog; future drawdowns may differ depending on the actual market path.

The full drawdown distribution (median, P25, P75, P95) is shown in the report’s “More details” expandable section.

Strategies typically don’t fail uniformly across markets — they have specific regime sensitivities. The site uses a 10-axis failure-mode taxonomy:

Each case in the catalog is tagged with one or more failure modes — Lehman 2008, for example, contributes to TREND_DOWN, SHARP_CRASH, and LIQUIDITY_STRESS simultaneously. The robustness score aggregates over all tagged contributions.

Operational definitions (drawdown thresholds, time scales, vol multipliers) and historical anchors per failure mode are documented on the Methodology page.

Two strategies tested on the same asset see the same case catalog and the same failure-mode taxonomy. Their robustness scores and per-regime breakdowns are directly comparable within this dataset.

Differences between strategies reflect different sensitivities to the specific cases and configurations tested. Results are conditional on the selected cases and configurations.

Cross-asset comparisons within a single strategy are valid but should account for asset-specific volatility baselines (e.g. equity vs. crypto). The site does not produce a single “best strategy” ranking — analysis is per-regime and per-asset.

Explicit limits on what the results can and cannot say:

• 01Results are conditional on the selected cases. The case catalog covers a curated set of historical and synthetic stress regimes — different cases would produce different scores.
• 02Synthetic stress probes are model-generated approximations, not statistical replicas of any real instrument. They are designed to isolate specific failure modes that historical data did not deliver cleanly.
• 03Volume-based indicators and conditions are not currently supported. OHLCV inputs include volume columns, but indicators such as OBV, VWAP, MFI, and CMF are not yet implemented in the engine — strategies referencing volume parse correctly, but volume-dependent signals do not fire.
• 04No forward-looking claims are made. The score reflects observed behavior in the tested cases; it is not a prediction of future performance.
• 05Strategy comparisons are valid within this dataset only. Differences between strategies reflect different sensitivities to the specific cases and configurations tested.
• 06Long-lookback strategies (e.g. SMA-200) are evaluated against synthetic cases via a 250-day antecedent pre-period that initializes indicator state before the stress phase begins. Cases where the warm-up requirement still exceeds the available pre-period plus case duration are excluded transparently in the report header.
• 07The robustness score is a relative measure of consistency across regimes — it does not measure absolute return potential or risk-adjusted return.