Strategy Lifecycle Retirement Criteria

When to Use

Use this skill when a multi-strategy book needs the retirement rule written down before the drawdown, and applied identically to every strategy. The failure mode it exists to prevent is not analytical — it is the researcher who built the strategy arguing that this quarter is unrepresentative, every quarter.

The engine measures nothing. It adjudicates: you supply already-computed live and backtest statistics, and it applies four pre-declared guardrails, a fixed escalation ladder, and produces a record of which criteria fired, which could not be evaluated, and which thresholds were in force. Feed it at your governance cadence (monthly or quarterly is typical), and route its output to the capital-allocation committee.

Its second job is to refuse. A payload with a NaN metric, a sign-inverted drawdown, or an unevaluable criterion does not come back ACTIVE_HEALTHY — it raises or downgrades. An engine that certifies a strategy on the absence of data is worse than no engine.

When NOT to Use

• On a short live track record. The engine cannot tell 10 live days from 10 years. An IR or an IC t-stat over a few weeks is noise, and retiring a good strategy on noise destroys capital as surely as never retiring a bad one. Supply live_observation_count and set min_live_observations; without both, no sample-size gate exists and the ladder will happily retire a three-week-old strategy.
• To diagnose why performance decayed. A market-wide regime shift and a strategy-specific alpha decay produce identical output here. Attribution belongs to strategy-performance-decay-detection-vs-market-wide-decay; without it, you will retire a perfectly good strategy in the middle of an unfavourable regime.
• As a live risk control. This runs on a governance cadence against aggregated statistics; it will not stop an intraday blow-up. The out-of-band control is kill-switch-and-drawdown-circuit-breakers and portfolio-level-stop-loss-independent-of-strategy-stops.
• As the unwind mechanism. MANDATORY_RETIREMENT is a decision, not an action. Blocking order entry and liquidating the book is strategy-decommissioning-and-position-unwind-procedure.
• With the default thresholds unexamined. All four are house heuristics. min_live_information_ratio=0.50 sits at the 75th percentile of Grinold & Kahn's active-manager table — the default retires anything below top-quartile. That may be exactly right for your book, or absurd; it is not a standard. See references/standards.md.

Prerequisites

• A StrategyPerformanceMetrics payload: strategy_id, backtest_sharpe, backtest_max_drawdown_pct, live_sharpe, live_max_drawdown_pct, live_information_ratio, live_ic_t_stat, live_realized_annual_return_pct, backtest_annual_return_pct, and optionally live_observation_count.
• Drawdowns as positive magnitudes. A 20% peak-to-trough decline is 20.0, never -20.0. Enforced, not inferred — see the pitfall below for why guessing is unsafe. backtest_max_drawdown_pct must be strictly positive; 0.0 is rejected as an unpopulated field.
• All metrics finite. Non-finite values raise; they do not degrade gracefully.
• An IR and an IC t-stat you can explain. The engine consumes both as opaque numbers. Whether the t-stat came from a one-sided or two-sided test, whether it was computed on overlapping windows without an autocorrelation adjustment, and how many observations sit behind it are all invisible here and all change what the 1.96 threshold means.
• backtest_sharpe and live_sharpe are carried as context only — no criterion reads them.
• Thresholds you are prepared to defend to a committee, and a change-control process around them: changing a retirement threshold is itself a material change under ESMA's supervisory expectations (references/standards.md).

Workflow

1. Validate the payload before adjudicating anything.

   • Reject non-finite metrics. Decision point — NaN is not a neutral value. nan < 0.50 is False, as is every other comparison, so a fully corrupt payload scores zero breaches and reads ACTIVE_HEALTHY. The engine raises instead.
   • Reject negative drawdowns and a zero backtested drawdown. Decision point — do not "helpfully" take the absolute value. Both sign conventions are common; guessing wrong is silent. Fail loudly and make the caller declare the convention.

2. Gate on sample size before running the ladder.

   • If min_live_observations is configured and the payload's live_observation_count falls short, return INSUFFICIENT_LIVE_HISTORY. Breaches are still computed and reported, for information only.
   • Decision point — a short track record is not evidence of decay, and it is not evidence of health either. The correct action is EXTEND_OBSERVATION at reduced size, not retirement and not full allocation.

3. Evaluate the four criteria. Comparisons are strict — a value sitting exactly on a threshold does not breach.

   • $\text{IR}_{\text{live}} < 0.50$ → ALPHA_DECAY_IR.
   • $\text{DD}{\text{live}} > 1.5 \times \text{DD}{\text{backtest}}$ → DRAWDOWN_BREACH.
   • IC t-stat $< 1.96$ → IC_STATISTICAL_DECAY.
   • $\text{Drift} = (R_{\text{live}} - R_{\text{backtest}}) / R_{\text{backtest}} \times 100 < -40%$ → PERFORMANCE_DRIFT.
   • Decision point — the drift ratio is not always computable. It is undefined when the backtested return is non-positive and numerically meaningless when it is near zero (0.1% backtest vs 0.05% live is a 5 bp miss that reads as $-50%$ drift). In both cases performance_drift_pct is None, the criterion is named in skipped_criteria, and return_gap_pct_points — the plain percentage-point difference, always defined — is reported instead. It is never substituted with 0.0.

4. Apply the escalation ladder.

   • 0 breaches, all four criteria evaluated → ACTIVE_HEALTHY.
   • 0 breaches, any criterion not evaluated → NEEDS_REVIEW. Decision point — ACTIVE_HEALTHY asserts that all four guardrails passed. A consumer reading only decision must never be told that on the strength of three.
   • 1 breach → NEEDS_REVIEW (watchlist).
   • 2 breaches → REDUCE_ALLOCATION (cut capital 50%, pending committee review).
   • $\ge 3$ breaches → MANDATORY_RETIREMENT.
   • Override — a drawdown breach with a live IR below zero retires at two breaches. A strategy simultaneously losing against its benchmark and exceeding its backtested worst case has failed on both axes that matter, and waiting for a third confirmation costs capital. The override is reported in escalation_reason, so a two-breach retirement is never unexplained. Disable it with escalate_on_negative_ir_with_drawdown_breach=False.

5. Archive the report, then hand off. thresholds_applied echoes every parameter in force — a threshold-dependent decision is not reproducible without it. MANDATORY_RETIREMENT is a decision; the actual order-entry block and position unwind are strategy-decommissioning-and-position-unwind-procedure.

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

Common Pitfalls

• Passing drawdowns as negative numbers. Under the negative convention, criterion 2 inverts: -30.0 > -15.0 is False, so a live drawdown three times the backtested worst case compares as within limits and the strategy is certified healthy. This is the highest-severity failure the skill has, and it is invisible — the report looks perfect. The engine now raises; do not "fix" it by taking abs() at the call site without first confirming which convention your data warehouse actually uses.
• Treating a missing measurement as a passing one. A drift of 0.0 returned because the backtested return was negative reads exactly like a strategy tracking its backtest perfectly. Any report where evaluated_criteria_count < 4 has an un-run guardrail, whatever breached_criteria says.
• Reading 1.96 as "95% confidence" in the decay question you are actually asking. 1.96 is the two-tailed 5% critical value of the standard normal; the one-sided 5% value is 1.645. "Has the IC decayed to non-positive?" is a one-sided question. Separately, an IC t-stat computed on overlapping forecast windows without a Newey-West correction is inflated, so the same 1.96 is a materially weaker test than it appears.
• Counting four breaches as four independent findings. IR, IC t-stat and return drift all degrade together when a signal stops working. "3 of 4 breached" is closer to one finding confirmed three ways. The ladder is a severity heuristic, not a statistical test, and the criteria are equally weighted by fiat.
• Retiring on a track record too short to mean anything. The most expensive misuse of this engine: a strategy three weeks into live trading breaches all four criteria on noise and gets decommissioned. If live_observation_count is None the gate silently does not run.
• Re-tuning the thresholds until the strategy passes. The engine's only real value is that the rule pre-dates the drawdown. Changing min_live_information_ratio from 0.50 to 0.30 because a favourite strategy is at 0.35 is the exact behaviour the skill exists to prevent — and under ESMA's supervisory briefing, changing risk-control thresholds is a material change requiring approval and a record.
• Emotional parameter tweaking on the strategy instead of the process. Repeatedly recalibrating a decaying strategy resets its live track record to zero each time, which conveniently makes every criterion unevaluable.
• Citing a regulator in support of these numbers. No regulator prescribes an IR floor, a drawdown multiple, or a t-stat cut-off for withdrawing a strategy. Retirement for economic underperformance is a business decision. What is regulated is different — see references/standards.md.

Verification

• Healthy baseline. StrategyLifecycleRetirementEngine() with IR $1.2$, IC t-stat $2.5$, live DD $11.0$ against backtest DD $10.0$, live return $18.0$ vs backtest $20.0$ → ACTIVE_HEALTHY, is_retired False, breached_criteria == [], evaluated_criteria_count == 4, performance_drift_pct == -10.0 (from $(18-20)/20 \times 100$), return_gap_pct_points == -2.0.
• Full decay. Backtest DD $8.0$ (allowed $12.0$), live DD $20.0$, IR $-0.2$, t-stat $0.4$, live return $-5.0$ vs backtest $25.0$ → MANDATORY_RETIREMENT, 4 breaches, drift $-120.0$.
• Boundaries — exactly on a threshold must pass. IR exactly $0.50$, t-stat exactly $1.96$, live DD exactly $15.0$ (= $1.5 \times 10.0$), and drift exactly $-40.0$ (live $12.0$ vs backtest $20.0$) all produce ACTIVE_HEALTHY. One tick past each must breach.
• Sign-convention regression. A live DD of $-30.0$ against a backtest DD of $-10.0$ must raise ValueError, not report ACTIVE_HEALTHY. A backtest_max_drawdown_pct of $0.0$ must raise.
• Corrupt-data regression. A NaN or Inf in any of the eight numeric fields must raise ValueError; a non-numeric value must raise TypeError.
• Unevaluable drift regression. Backtest return $-3.0$ with live return $-40.0$ → performance_drift_pct is None, evaluated_criteria_count == 3, NEEDS_REVIEW (not ACTIVE_HEALTHY), return_gap_pct_points == -37.0. Backtest $0.1$ vs live $0.05$ → drift None and no manufactured breach. Backtest exactly $1.0$ vs live $0.5$ → drift $-50.0$, evaluated normally.
• Escalation override. Live DD $20.0$ with IR $-0.2$ → MANDATORY_RETIREMENT at 2 breaches, escalation_reason containing OVERRIDE_DD_AND_NEGATIVE_IR. With escalate_on_negative_ir_with_drawdown_breach=False the same payload → REDUCE_ALLOCATION.
• Sample-size gate. With min_live_observations=126, a payload breaching all four criteria at live_observation_count=15 → INSUFFICIENT_LIVE_HISTORY, is_retired False, four breaches still listed. At $126$ observations the same payload → MANDATORY_RETIREMENT.
• Constructor validation. A non-positive max_drawdown_multiplier, a non-finite threshold, a mandatory_retirement_breach_count outside $1..4$, and a non-positive min_live_observations must each raise ValueError.
• Run python -m unittest discover -s skills/strategy-lifecycle-retirement-criteria/scripts — 45 tests, 100% pass rate.

Related Skills

• strategy-performance-decay-detection-vs-market-wide-decay
• strategy-decommissioning-and-position-unwind-procedure
• strategy-underperformance-remediation-decision-tree
• strategy-committee-governance-for-capital-allocation-decisions
• backtest-vs-live-performance-divergence-tracking
• capital-reallocation-based-on-live-performance