Strategy Underperformance Remediation Decision Tree

When to Use

Use this skill when a live strategy has underperformed and someone is about to do something about it. The failure mode it exists to prevent is not analytical — it is the researcher who reaches for the parameter file first, every time, whatever the cause.

One symptom, a fallen Sharpe ratio, has four incompatible causes, and each remedy is wrong for the other three. Recalibrating parameters when the real problem is a stale data feed fits the model to the defect. Decommissioning during a market-wide regime shift permanently forfeits the recovery. Tuning the signal when 75% of the alpha is being eaten by slippage optimizes the wrong system entirely. The engine applies a pre-declared triage order to a metrics payload and returns an auditable routing decision plus the record of which nodes were tested on the way there:

Its second job is to refuse. A payload with a NaN Sharpe, a sign-inverted slippage figure, a stringified boolean, or a zero expected alpha does not come back MAINTAIN_TRADING — it raises. An engine that certifies a strategy healthy on the absence of usable data is worse than no engine.

When NOT to Use

• As a significance test of strategy versus peers. Nodes 3 and 4 are two independent threshold comparisons, not a hypothesis test, and the split between them has no null distribution. For a statistic that does — the Memmel-corrected Jobson-Korkie Sharpe-difference test — use strategy-performance-decay-detection-vs-market-wide-decay. This engine routes; that one tests.
• On a short live track record, ungated. Set min_live_observations or the tree will happily route a three-week-old strategy to decommissioning. At 60 daily observations the standard error of an annualized Sharpe near 1.0 is $\approx 2.05$ (Lo 2002) — larger than the threshold being tested. Check sharpe_evidence_conclusive before writing "underperforming" in a committee memo; it will usually be False.
• As the executor of its own recommendation. Every outcome is a field on a dataclass. Nothing here cancels an order, liquidates a position, halts a strategy, changes an allocation, or re-optimizes anything. Unwinding is strategy-decommissioning-and-position-unwind-procedure; halting is kill-switch-and-drawdown-circuit-breakers, which must remain structurally independent of strategy logic; the capital change is capital-reallocation-based-on-live-performance.
• As a live risk control. This runs at a governance cadence against aggregated statistics. It will not stop an intraday blow-up, and Node 3's "cut 50%" is a committee recommendation, not a circuit breaker.
• Without a defensible peer benchmark. The Node 3 / Node 4 split is only as good as the peer index. Benchmarking a market-neutral book against a long-only index manufactures a "parameter drift" verdict every time the index rallies. Choosing the benchmark is benchmark-selection-for-strategy-evaluation; this engine only compares against the number you supply.
• To decide the retirement rule itself. The four criteria that constitute "this strategy has failed" belong in strategy-lifecycle-retirement-criteria, declared before the drawdown. This engine assumes the underperformance is already established and asks only why.

Prerequisites

• An UnderperformanceTriageMetrics payload: strategy_id, live_sharpe, backtest_sharpe, peer_benchmark_sharpe, realized_slippage_bps, expected_alpha_bps, is_data_feed_healthy, is_alpha_hypothesis_valid, and optionally live_observation_count.
• Five caller conventions the engine cannot verify. Getting any of them wrong silently changes the routing:
  • is_alpha_hypothesis_valid is a human research judgment about the economic mechanism — has the inefficiency been arbitraged away, has the market microstructure that created it changed — and not a re-reading of live P&L. Derive it from recent returns and Node 1 becomes a Sharpe threshold with an irreversible consequence, firing during exactly the regime shifts Node 3 exists to survive.
  • realized_slippage_bps is a positive cost magnitude in basis points. A signed convention (cost as a negative number) is rejected, not reinterpreted — see the pitfalls for why guessing is unsafe. Pass 0.0 for net price improvement.
  • realized_slippage_bps and expected_alpha_bps must share a horizon and a unit of activity — both per round trip, or both per unit time. Per-trade slippage against annualized alpha understates the ratio by roughly the number of trades per year and defeats Node 2 entirely.
  • All three Sharpe ratios share an annualization convention and return frequency, or the comparisons are meaningless.
  • live_observation_count counts the live return observations behind live_sharpe, not backtest observations.
• backtest_sharpe is carried as context only — no node reads it. The report returns sharpe_gap_vs_backtest for the committee's benefit.
• Thresholds you are prepared to defend. min_healthy_sharpe = 1.0 and min_peer_sharpe = 0.50 are house defaults, not external standards — see references/standards.md.
• A change-control process downstream. Acting on RECALIBRATE_MODEL_PARAMETERS is a material change requiring authorisation, recording, and retesting (references/standards.md).

Workflow

1. Validate the payload before routing anything.

   • Decision point — NaN is not a neutral value, it is the most dangerous one. Every node is a threshold comparison, and every comparison against NaN is False. A NaN live_sharpe therefore fails live_sharpe < 1.0 at Nodes 3 and 4 and falls straight through to the healthy branch, so a fully corrupt payload returns MAINTAIN_TRADING — "no remediation required". The engine raises ValueError instead.
   • Decision point — reject a non-positive expected_alpha_bps, do not divide by it. The slippage ratio is undefined at zero and sign-inverted below it, so Node 2 would clear however large the slippage was. A strategy with no expected alpha has no edge to protect: that is a Node 1 hypothesis question, not a Node 2 execution question.
   • Decision point — reject a stringified boolean. is_alpha_hypothesis_valid="False" is truthy. A JSON payload or an agent that passes the string clears Node 1 and keeps a dead strategy trading.

2. Node 1 — Fundamental hypothesis audit (resolves first, outranks everything):

   • Hypothesis invalid $\implies$ MANDATORY_STRATEGY_DECOMMISSION. It outranks a perfect Sharpe and flawless execution, because there is nothing to recalibrate and no execution fix that restores an edge that no longer exists.
   • Decision point — two situations make this judgment suspect, and the report says so rather than blocking it. If the data feed is flagged unhealthy, the metrics the judgment was formed against may be corrupt, and decommissioning is irreversible where a feed repair is not. If the peer group is also impaired, the "decay" may be the regime shift Node 3 exists to survive. Both surface in warnings.

3. Node 2 — Execution and data quality audit (must precede the Sharpe nodes):

   • Feed unhealthy, or $\frac{\text{slippage}}{\text{alpha}} > 50%$ (strict: exactly $50%$ clears) $\implies$ OPTIMIZE_EXECUTION_AND_DATA.
   • Decision point — this node comes before Nodes 3 and 4 for a reason. A broken feed or a cost that swallows the alpha invalidates the very Sharpe ratios Nodes 3 and 4 compare. Tuning signal parameters against fills destroyed by slippage fits the model to the execution defect and buries it.
   • When the feed is unhealthy, warnings records that every Sharpe figure in the report came from a feed the caller flagged as unreliable. Re-run the triage after the repair before acting on any comparison.

4. Node 2a — Sample-size gate (only when min_live_observations is configured):

   • Live window shorter than the minimum $\implies$ EXTEND_OBSERVATION_INSUFFICIENT_HISTORY, at reduced size.
   • Decision point — a short track record is not evidence of decay, and it is not evidence of health either. It sits below Nodes 1 and 2 deliberately: a dead hypothesis and a broken feed are not sample-size questions, and a young strategy should not be shielded from either.
   • Configuring the gate while omitting live_observation_count raises rather than silently ungating.

5. Node 3 — Market-wide regime shift audit (resolves before Node 4):

   • Strategy below mandate and peer benchmark below its health floor $\implies$ TEMPORARY_CAPITAL_DEGRADE_RETAIN_SIGNAL (cut allocation 50%, retain the signal).
   • Decision point — joint impairment resolves first because it is the conservative reading of identical evidence. When peers are down too, the data does not support an idiosyncratic verdict, and a capital cut is reversible where a recalibration that overwrites the parameter set is not.

6. Node 4 — Parameter drift audit:

   • Strategy below mandate while peers are healthy $\implies$ RECALIBRATE_MODEL_PARAMETERS.
   • Decision point — check sharpe_gap_vs_peer before re-optimizing. The two thresholds are asymmetric (mandate $1.0$, peer floor $0.50$), so a strategy at $0.95$ against peers at $0.55$ lands here while outperforming its cohort. It is missing its own mandate without underperforming its peers, which is weak evidence of drift; the report warns, and the right response is usually to review the mandate or the benchmark, not the parameters.

7. Read warnings, then route the report to governance. An empty tuple means no caveats applied. Every Sharpe-driven outcome carries an evidence caveat when the sample size is unknown or the estimate sits within 1.96 standard errors of the mandate threshold, and every recalibration carries the material-change caveat. The report is a recommendation for authorisation, not an executed decision.

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

Common Pitfalls

• A NaN metric certifying a strategy healthy: float('nan') < 1.0 is False, so a corrupt payload clears Nodes 3 and 4 and returns MAINTAIN_TRADING. The most permissive verdict the engine can produce, issued on unusable data. Reject non-finite values; never let them reach a comparison.
• A signed slippage convention disarming the execution node: if your TCA system reports cost as $-15$ bps, the ratio is $-0.75$, which never exceeds $0.50$, so a strategy losing three quarters of its alpha to execution is routed to RECALIBRATE_MODEL_PARAMETERS instead. The convention is a positive magnitude, and violating it is silent.
• Zero or negative expected alpha bypassing the ratio test: guarding the division with expected_alpha_bps > 0 and skipping the node when it fails means the strategies with no measurable edge — the ones most in need of the execution audit — are the ones that skip it.
• Comparing per-trade slippage against annualized alpha: a strategy paying 15 bps per round trip against 300 bps of annual alpha scores a 5% ratio and clears Node 2, when the horizon-matched figure at 100 trades a year is 500%.
• HARKing (Hypothesizing After Results are Known): bolting an indicator onto a strategy to "fix" it, without establishing whether execution slippage, a regime shift, or alpha decay was the cause. The node order exists precisely to stop the parameter file being the first reflex.
• Deriving the alpha hypothesis from live P&L: it turns Node 1 into a Sharpe threshold whose consequence is irreversible, and it fires hardest during the drawdowns Node 3 is designed to survive. The judgment is about the economic mechanism, made by research, ideally recorded before the drawdown.
• Decommissioning during a regime shift: retiring a sound strategy while its whole peer group is equally impaired destroys capacity that cannot be rebuilt when the regime turns.
• Recalibrating a strategy that is beating its peers: with a $1.0$ mandate and a $0.50$ peer floor, a strategy at $0.95$ against peers at $0.55$ is routed to recalibration while leading its cohort. Read sharpe_gap_vs_peer, not just the action.
• Reading a routing decision as a demonstrated finding: at 60 daily observations the standard error of an annualized Sharpe near 1.0 is $\approx 2.05$. A strategy observed at 0.6 against a 1.0 mandate is not shown to be underperforming. The gates are policy floors, not statistical conclusions.
• Letting recalibrations accumulate untested: ESMA's February 2026 supervisory briefing (¶30) warns that "a series of minor or small changes due to recalibrations could accumulate over time, when uncontrolled or unchecked, into a material change in the model output without it being tested". A remediation engine emitting RECALIBRATE_MODEL_PARAMETERS on a monthly cadence is exactly that mechanism — record and retest each one.
• Treating the report as the action: is_decommissioned=True sets a boolean. Something else must cancel the orders and liquidate the book.

Verification

• Instantiate StrategyUnderperformanceRemediationEngine. Invalid alpha hypothesis with an otherwise perfect payload (Sharpe 2.5, healthy peers, 7% slippage ratio) $\implies$ MANDATORY_STRATEGY_DECOMMISSION, is_decommissioned=True, decisive_node="NODE_1_HYPOTHESIS_FAILURE". 15 bps slippage against 20 bps alpha (75%) $\implies$ OPTIMIZE_EXECUTION_AND_DATA. Unhealthy feed with 0.0 bps slippage $\implies$ OPTIMIZE_EXECUTION_AND_DATA and a caveat that every Sharpe in the report came from the bad feed. Strategy 0.6 against peers 1.5 $\implies$ RECALIBRATE_MODEL_PARAMETERS, is_capital_reduced=False. Strategy 0.3 against peers 0.2 $\implies$ TEMPORARY_CAPITAL_DEGRADE_RETAIN_SIGNAL.
• Verify annualized_sharpe_standard_error(1.50, 60, periods_per_year=1) == 0.188 and (3.00, 60, periods_per_year=1) == 0.303, reproducing Lo (2002) Table 1, and that (1.0, 60) == 2.0514223 and (0.0, 2520) == 0.3162278 at the 252-day default.
• Boundary checks: slippage exactly at $50%$ clears Node 2 (strict >), $50.05%$ fires it; live_sharpe exactly $1.0$ is healthy (inclusive), $0.9999$ is not; peer_benchmark_sharpe exactly $0.50$ selects Node 4, $0.4999$ selects Node 3.
• Negative checks — each must raise ValueError, not route: NaN/±Inf on any of the five numeric fields; a negative realized_slippage_bps; expected_alpha_bps of 0.0 or -5.0; "False", "true", 0, 1 or None for either boolean flag; an empty or non-string strategy_id; a non-positive or non-integer live_observation_count; min_peer_sharpe > min_healthy_sharpe; a non-positive max_slippage_alpha_ratio or periods_per_year; and a configured min_live_observations with no live_observation_count in the payload.
• Confirm warnings fires on: a Node 1 judgment made against an unhealthy feed, a Node 1 judgment made while peers are impaired, a Node 4 routing where sharpe_gap_vs_peer >= 0, a healthy strategy against an impaired peer group, any Sharpe-driven routing with no live_observation_count, and every recalibration. Confirm a healthy strategy with 5,000 conclusive observations returns warnings == ().
• Confirm triage_path records the cleared nodes in order with the decisive node last, and that unreached nodes are absent entirely — a cleared node in the record is evidence it was tested.
• Run python -m unittest discover -s skills/strategy-underperformance-remediation-decision-tree/scripts and confirm 100% pass rate (43 tests).

Related Skills

• strategy-performance-decay-detection-vs-market-wide-decay
• strategy-lifecycle-retirement-criteria
• strategy-decommissioning-and-position-unwind-procedure
• benchmark-selection-for-strategy-evaluation
• transaction-cost-analysis-tca-integration
• backtest-vs-live-performance-divergence-tracking
• capital-reallocation-based-on-live-performance
• kill-switch-and-drawdown-circuit-breakers