Pivot Techniques
Three systematic techniques for generating structurally different strategy proposals when the backtest iteration loop stalls.
Technique 1: Assumption Inversion
Principle: Identify the core assumptions of the current strategy and invert them based on which stagnation trigger fired.
Inversion Rules by Trigger
| Trigger | Module | Inversion |
|---|---|---|
cost_defeat |
horizon | Shorten holding period (reduce friction exposure) |
cost_defeat |
universe | Shift to higher-liquidity names (reduce slippage) |
tail_risk |
risk | Tighten stops, reduce position size, add max drawdown cap |
tail_risk |
structure | Move toward market-neutral or hedged approach |
improvement_plateau |
signal | Change signal source (price-based → volume/fundamental) |
improvement_plateau |
entry | Change entry timing mechanism |
overfitting_proxy |
complexity | Reduce parameters, use simpler entry model |
overfitting_proxy |
validation | Extend test period, add regime subsamples |
Application Process
- Read the trigger(s) fired from the diagnosis
- Look up the corresponding inversion rules
- Apply each inversion to the current strategy's modules
- Generate a new draft with the inverted assumptions
Technique 2: Archetype Switch
Principle: Jump to a structurally different strategy archetype that addresses the same market inefficiency from a different angle.
Process
- Identify the current strategy's archetype from its
hypothesis_type,mechanism_tag, andentry_family - Look up compatible pivot targets in the archetype catalog's
compatible_pivots_frommapping - For each compatible target, generate a draft using the target archetype's default modules
- Preserve the original concept_id and thesis where applicable
Archetype Identification Rules
| hypothesis_type | mechanism_tag | entry_family | Archetype |
|---|---|---|---|
| breakout | behavior | pivot_breakout | trend_following_breakout |
| breakout | structural | pivot_breakout | volatility_contraction |
| mean_reversion | statistical | * | mean_reversion_pullback OR statistical_pairs |
| mean_reversion | information | * | event_driven_fade |
| earnings_drift | information | gap_up_continuation | earnings_drift_pead |
| momentum | behavior | * | sector_rotation_momentum |
| regime | macro | * | regime_conditional_carry |
When multiple archetypes match, prefer the one whose entry_family matches the current draft.
Technique 3: Objective Reframe
Principle: Change what "success" means for the strategy, shifting the optimization target.
Reframe Options
| Current Objective | Reframed To | Rationale |
|---|---|---|
| Maximize Sharpe ratio | Minimize max drawdown | Better tail risk control |
| Maximize expectancy | Maximize win rate | More consistent (smaller but more frequent wins) |
| Maximize total return | Maximize risk-adjusted return per unit exposure | Capital efficiency focus |
Application
- Read the current strategy's
success_criteriafromvalidation_plan - Select a reframe based on the trigger:
tail_risk→ drawdown minimizationcost_defeat→ win rate maximization (smaller targets, tighter stops)improvement_plateau→ risk-adjusted return (different efficiency lens)overfitting_proxy→ simpler criteria (fewer optimization targets)
- Adjust exit rules and risk parameters to align with the new objective
- Update
success_criteriain the generated draft
Technique Selection by Trigger
| Trigger | Primary Technique | Secondary Technique |
|---|---|---|
improvement_plateau |
Archetype Switch | Assumption Inversion |
overfitting_proxy |
Assumption Inversion | Objective Reframe |
cost_defeat |
Assumption Inversion | Archetype Switch |
tail_risk |
Assumption Inversion | Objective Reframe |
The generate_pivots script applies all applicable techniques and lets scoring determine the best candidates.
Scoring
Quality Potential (0-1)
Heuristic score based on how well a target archetype addresses the specific trigger. Defined in QUALITY_TABLE dictionary mapping (trigger, archetype) → score.
Novelty (0-1)
Jaccard distance between the module sets of the original strategy and the proposed pivot:
novelty = 1 - |A ∩ B| / |A ∪ B|Where A and B are sets of (key, value) pairs constructed from:
("hypothesis_type", <value>)("mechanism_tag", <value>)("regime", <value>)("entry_family", <value>)("horizon", <"short"|"medium"|"long">)— time_stop_days ≤ 7: short, ≤ 30: medium, else: long("risk_style", <"tight"|"normal"|"wide">)— stop_loss_pct ≤ 0.04: tight, ≤ 0.08: normal, else: wide
Combined Score
combined = 0.6 * quality_potential + 0.4 * noveltyTiebreak Rules (deterministic)
- Combined score descending
- Novelty descending (prefer more novel proposals)
- Proposal ID alphabetical ascending (deterministic final sort)
Diversity Constraint
Maximum 1 proposal per target archetype. If multiple techniques produce candidates for the same archetype, keep the one with the highest combined score.