Multi Year Regime Coverage Requirement

When to Use

Invoke this skill during strategy validation, before promotion to live capital. A strategy with stellar metrics over one or two years of a trending bull market routinely suffers a catastrophic drawdown in the first high-volatility crash or sideways range it meets. This skill buckets historical bars into market regimes, enforces a minimum coverage floor (duration and at least min_required_regimes regimes, each with enough bars to be more than an artifact), and reports drawdown, return, win rate and Sharpe per regime so the worst bucket is visible rather than averaged into an aggregate.

Use it as a gate: an aggregate multi-year Sharpe of 2.5 tells you nothing about whether the strategy lost 40% during the one crash in the sample.

When NOT to Use

• As a live regime signal. Bar i's label is computed from a window that includes bar i's own price. That is correct for retrospective attribution and catastrophic as a trading signal — at the open of bar i that price does not exist yet. For live routing use regime-detection-for-strategy-switching, which confirms regimes with hysteresis on already-closed bars.
• On intraday bars while leaving bars_per_year at 252. The default is daily. One year of 1-minute bars is 98,280 bars, which at 252 audits as 390 years of coverage and passes any duration gate. Frequency cannot be inferred from a list of floats — you must set it.
• On a series with calendar gaps. Duration is bar count ÷ bars_per_year, not a timestamp span. A series missing six months reports the span it would have had if the bars were consecutive.
• As the account-level drawdown control. This measures drawdown inside each regime bucket. The portfolio-level question belongs to kill-switch-and-drawdown-circuit-breakers and portfolio-level-stop-loss-independent-of-strategy-stops.
• As proof of robustness on its own. Passing is necessary, not sufficient. It says nothing about look-ahead bias (lookahead-bias-elimination), overfitting (walk-forward-validation-setup), or capacity.

Prerequisites

• A gap-free price series, strictly positive and finite, at one consistent bar frequency.
• bars_per_year matching that frequency (252 for daily bars) — it drives both the duration gate and the Sharpe annualization.
• A strategy return series aligned one-to-one with the prices: strategy_returns[i] is the return realized over the bar ending at prices[i]. Same length, no exceptions — the engine raises rather than truncating.
• Returns expressed as differential (excess) returns if the risk-free rate is non-zero; otherwise the reported figure is a return-to-variability ratio, not a Sharpe ratio.
• Thresholds you are willing to defend: minimum years, minimum regimes, minimum bars per regime, and the per-regime drawdown limit. The defaults are house heuristics, not standards — see references/standards.md.

Workflow

1. Segment Historical Data into Market Regimes:

   • For each bar with a full trailing window, compute annualized volatility $\sigma = \text{stdev}(r_{\text{window}}) \times \sqrt{F}$ and window price change $\Delta = (P_i - P_{i-w}) / P_{i-w}$.
   • Bucket: $\sigma >$ high_vol_annualized_threshold → HIGH_VOLATILITY_CRASH; else $\Delta > $ trend_threshold_pct → BULL_TREND; else $\Delta < -$trend_threshold_pct → BEAR_MARKET; else LOW_VOLATILITY_RANGE.
   • Decision point — warm-up bars are UNCLASSIFIED, not "range". The first $w$ bars have no trailing window. Labelling them LOW_VOLATILITY_RANGE invents a regime that then satisfies part of the coverage requirement. They are excluded from coverage and from every metric, and counted in unclassified_bars so the exclusion is auditable.

2. Audit Multi-Year Regime Coverage:

   • Duration: $\text{bars} / F \ge T_{\text{min}}$, compared unrounded — 755 daily bars is 2.9960 years and must not pass a 3-year gate that needs 756.
   • Decision point — a regime only counts if it has at least min_bars_per_regime bars. Otherwise a 750-bar bull backtest with one incidental bear bar and one crash bar "covers three regimes". Thin regimes are still reported in regimes_observed and still carry metrics; they just do not satisfy the coverage floor.

3. De-average Performance Across Regimes:

   • Per regime: compounded total return, win rate, annualized Sharpe, and drawdown.
   • Decision point — Sharpe is None, never a number, when it is not measurable. Below min_bars_per_regime observations, or when return dispersion is floating-point noise (a constant return series), there is no risk-adjusted interpretation. Substituting an epsilon denominator produced a reported Sharpe of $2.4 \times 10^{16}$; None means "not measurable" and must be rendered as such, not coerced to 0.
   • Decision point — two drawdown figures, only one of them real. max_drawdown_pct is the worst decline within a single contiguous episode of that regime: the decline an account actually experienced, and the veto metric. concatenated_drawdown_pct chains every bar of the regime together, skipping the bars in between — useful for spotting death by a thousand cuts across many separate episodes, but it is a decline that never occurred. Never report it as realized.

4. Enforce Promotion Thresholds:

   • Veto promotion if within-episode drawdown exceeds max_allowed_regime_drawdown_pct in any regime, however high the aggregate Sharpe.
   • Decision point — a thin regime still vetoes. A 40% loss over five crash bars is a path fact, not a statistical estimate; small sample size disqualifies a Sharpe ratio, not a drawdown.
   • Decision point — report each failure for what it is. A coverage failure and a drawdown breach are different findings. vetoed_regimes is empty when no drawdown breach occurred, and the message must not claim one.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Reusing the classifier as a live signal: the label for bar i is built from a window ending at bar i. A backtest that routes strategies on these labels knows the current bar's close before the bar opens and will look spectacular for reasons that have nothing to do with the strategy.
• Counting a regime that barely appeared: without a minimum bar count, "covers 3 regimes" is satisfiable by two stray bars in an otherwise single-regime sample. The gate then certifies exactly the backtest it exists to reject.
• Feeding intraday bars at the daily default: 98,280 one-minute bars audited at 252 bars/year report 390 years of coverage. The duration gate passes trivially and the annualized Sharpe is overstated by roughly $\sqrt{390}$.
• Letting corrupt data pass the gate: a single NaN return makes the compounded equity curve NaN, and every subsequent drawdown > max_drawdown comparison is False. Max drawdown stays 0.00%, no veto fires, and the strategy is certified by the absence of data. Reject non-finite prices and returns before auditing — the engine raises.
• Silently misaligned series: a return series one element shorter than the price series is the most common off-by-one in backtesting. Truncating to the shorter series shifts every regime label by one bar; the engine raises instead.
• Testing only in bull regimes: backtesting momentum strategies exclusively across the 2020–2021 liquidity expansion, masking the 2022 drawdown.
• Reporting a single aggregate Sharpe: hiding a $-40%$ crash drawdown behind an overall $2.5$ Sharpe driven by one favourable regime.
• Arbitrary window slicing: hand-picking start/end dates that exclude crash periods instead of classifying algorithmically.
• Treating the labels as conventional market definitions: BEAR_MARKET here means a 20-bar move below $-3%$. The SEC's investor education material defines a bear market as a broad index falling 20% or more over at least two months. Do not write "tested through a bear market" in an external document on the strength of this bucket.
• Comparing regime Sharpe ratios across audits with different thresholds: the buckets are threshold-dependent. Record the parameters with the report or the numbers are not reproducible.

Verification

• Instantiate MarketRegimeCoverageEngine(min_required_years=1.0, min_required_regimes=3, max_allowed_regime_drawdown_pct=25.0) and feed the 540-bar four-regime fixture from scripts/test_regime_coverage.py. Confirm 20 UNCLASSIFIED warm-up bars, all four regimes covered, and total_years = 540/252 = 2.1429.
• Place two $-10%$ bars at the start of each of the two LOW_VOLATILITY_RANGE episodes. Verify max_drawdown_pct = 19.00 (within-episode: $1.0 \to 0.90 \to 0.81$), concatenated_drawdown_pct = 34.39 ($0.9^4 = 0.6561$), total_return_pct = $-34.39$ (compounded, not the arithmetic sum of $-40.00$), and that no veto fires.
• Place six consecutive $-5%$ bars inside BULL_TREND: $0.95^6 = 0.735092$, a 26.49% decline. Verify is_coverage_sufficient is True and is_promotable is False, so the veto is demonstrably the only reason for refusal.
• Boundary checks: a $-25.0049%$ bar must veto even though it reports as 25.00%; exactly $-25%$ must not. 755 daily bars must fail a 3-year gate; 756 must pass.
• Negative checks: a NaN return, a NaN or non-positive price, a return at or below $-100%$, a length mismatch, and each out-of-range constructor argument must all raise ValueError.
• Confirm a constant return series reports sharpe_ratio is None for every regime, not a large number.
• Run python -m unittest discover -s skills/multi-year-regime-coverage-requirement/scripts and confirm 100% pass rate.

Related Skills

• walk-forward-validation-setup
• regime-detection-for-strategy-switching
• stress-testing-against-historical-crash-scenarios
• monte-carlo-strategy-robustness-testing
• paper-to-live-promotion-checklist