Skip to content

strategy-optimization.md

Latest commit

 

History

History
236 lines (177 loc) · 12.7 KB

strategy-optimization.md

File metadata and controls

236 lines (177 loc) · 12.7 KB
Raw

Strategy Optimization Research — Overnight Earnings (Defined-Risk)

Research-backed optimization hypotheses for the seven remaining defined-risk strategies (iron_fly, iron_condor, double_calendar, atm_calendar, directional_credit_spread, broken_wing_butterfly, reverse_fly), specific to this system's structure: one position opened before the close, closed once after the next open, unmonitored overnight — so the entire edge is the single overnight IV-crush event, and strike/structure choices at entry matter more than intraday management (there is none).

These are hypotheses to validate, not changes to apply. The right way to act on them is the existing paper-testing program (docs/strategy-testing-plan.md): strategy_test_runner already force-samples all 7 strategies nightly, and strategy_metrics already computes cost-adjusted expectancy and captured IV crush per strategy. Change one parameter, re-run the paper program, compare expectancy — do not blind-tune to the literature. The research itself warns that "if all your backtest metrics are excellent, you probably have overfitting."

Every current config value cited below is from config/config.json at time of writing.

0. The removal itself is the #1 research-backed improvement

The strongest, most consistent caution in the earnings-options literature is against undefined risk on single-name earnings events — a naked short can gap arbitrarily. Making the whole system defined-risk-only (max loss known at entry for every trade) directly implements that. This is already done; everything below optimizes within defined risk.

1. Profit target for IV-crush plays

Research: for earnings IV-crush plays, take profits early — commonly 50-75% of max (vs. the ~50% standard for non-event condors), because IV crush deflates the position in one move and holding past that only re-exposes you to gamma/gap risk.

Current values: iron_fly / iron_condor / directional_credit_spread / reverse_fly = 0.50; atm_calendar / double_calendar = 0.25; broken_wing_butterfly = 0.10.

Hypotheses:

  • BWB 0.10 looks far too low — a 10% profit target on an earnings crush leaves most of the captured move on the table. Test raising toward 0.25-0.50.
  • Calendars at 0.25 and credit spreads at 0.50 sit at/below the earnings 50-75% band — test upward sensitivity (e.g. 0.60-0.75 on the credit spreads).
  • Nuance specific to this system: the overnight hold captures the crush in a single move and the close-window backstop force-exits regardless, so the profit target mostly governs the Step-3c early-exit check the next morning, not a mid-hold decision. Its practical effect is smaller here than in a multi-day condor — measure whether it moves expectancy at all before tuning it.

Metric to watch: per-strategy expectancy and win rate vs. target; cross-reference iv_crush (captured vol points) to confirm the crush actually happened on winners.

2. Iron Fly (ATM) vs Iron Condor (OTM) on an unmonitored hold

Research: short-strike delta sets probability of profit — ~16-delta shorts ≈ 68% POP (conservative), 30-40-delta for more premium. An iron fly sells ATM = maximum premium but lowest POP and highest pin/gamma risk; an iron condor sells OTM = wider profit zone, higher POP, smaller credit.

Current: iron_fly sells the ATM straddle (capped at max_atm_delta_abs 0.57); iron_condor places shorts at the expected-move boundary.

Hypothesis: on an unmonitored overnight hold, the ATM fly's pin/gamma risk is the worst profile (nobody adjusts if the stock lands near the short strike), so iron_condor should out-expectancy iron_fly despite the smaller credit. The paper program already trades both head-to-head every night — let the data rank them rather than assuming the fly's larger credit wins.

3. Skew-based strike/wing adjustment for fly, butterfly, condor

Research: symmetric strikes are often suboptimal. Put-side skew makes downside strikes richer (more premium) and makes their delta overstate real risk, so mirroring wings around spot leaves edge on the table. The fix is asymmetric (broken) wings / asymmetric short-strike deltas that lean with the skew — push/widen the richer (usually put) side. Broken-wing structures are the tool research most recommends for skewed earnings names.

Current state is split:

  • broken_wing_butterfly and directional_credit_spread already consume skew (min_skew_abs 0.02, skew_delta_target 0.25, asymmetric near/far wings, put-vs-call side chosen by the skew sign).
  • iron_fly (symmetric ATM straddle + equal wings via wing_width_credit_multiple) and iron_condor (symmetric expected-move boundaries) ignore skew entirely.

The skew signal is already computed in the shared engine (scanner's call/put ATM-IV skew, consumed by BWB/directional today), so it can be reused with no new data path.

Hypotheses:

  1. Add optional skew-leaned wings to iron_fly / iron_condor — widen or push the richer-skew wing (usually the put side) and/or offset the short strikes toward the skew — and compare cost-adjusted expectancy vs. today's symmetric build head-to-head.
  2. Validate BWB's existing skew logic: confirm the side selection and skew_delta_target actually place the fat wing on the correct (richer-IV) side on real candidates.

If the paper data supports (1), the follow-up is a small skew-offset parameter on the two symmetric strategies, mirroring how BWB/directional already do it — doc/hypothesis only in this pass, no strike-construction code changes yet.

4. Calendar term-structure / DTE

Research: the calendar edge is front-month IV ≫ back-month IV (steep backwardation into earnings), with the front leg expiring ~2-3 days after the report — enough to catch the front-leg crush while the back leg keeps its value. The back leg also loses value in a crush, so profit is the net differential, not just front-leg decay.

Current: atm_calendar / double_calendar use back_month_min_days_after 21 and pick the front expiration near the reaction date.

Hypothesis: confirm the selected front expiration lands a few days after earnings (not a same-day 0DTE, not weeks out). The IV/RV + term-structure gate already encodes "front IV elevated"; the open question is purely the front-leg DTE placement. Instrument it in the paper entry_context and check the realized front-vs-back IV differential on winners.

5. Expected move > realized move (the core edge)

Research: prioritize names where the expected (priced-in) move exceeds the historical realized move — options are overpriced, so the crush works in your favor. This is the single most-cited edge for short-premium earnings trades.

Current: the agent already gates on IV/RV ratio (min_iv_rv_ratio 1.25) and realized-move dispersion (realized_move_dispersion_pct) — a direct implementation of "expected exceeds realized, consistently."

Hypothesis: this is the agent's primary edge and is already well-aligned with the research — the lever is to tighten rather than loosen it. Test whether raising the IV/RV floor (e.g. 1.25 → 1.35) improves per-trade expectancy at the cost of fewer trades.

6. Loss limit (stop)

Research: a common stop is 2-3× credit collected ("don't hope and pray").

Current: stop_loss_credit_multiple 1.5 on the credit strategies (tighter than the research norm); calendars use stop_loss_pct_of_debit 1.0; reverse_fly stops at its defined max loss.

Hypothesis: because the stop is only checked at the open (unmonitored overnight), a tight 1.5× may cut positions that would have recovered intraday, while a looser 2-3× may just book bigger gap losses. Test 1.5× vs 2.0× and measure the effect on the loss tail (avg loss, max drawdown) — this is a where-does-it-actually-fire question the paper data answers.

7. Wing width (low priority)

Research: wing width is a risk/return trade-off (wider = lower % return but more buffer); $5-10 on stocks is typical, and a credit-multiple approach is fine.

Current: credit-multiple wings (2.5-3.5× bands). Consistent with the research framing — low priority, revisit only if a strategy underperforms after the higher-priority levers.

8. ATM bid-ask spread gate (max_bid_ask_spread_pct) — flagged for review

Observation (2026-07-09): DAL — a $61B-market-cap, high-volume name, otherwise clean on every other criterion (term structure, expected move, delta, OI, weekly options) — was hard-rejected across all 7 strategies for bid_ask_spread_too_wide: measured ATM spread 16.47% vs. the max_bid_ask_spread_pct 0.15 cap, a ~1.5-point miss. Single data point, not acted on yet — this section exists to track whether it recurs.

Current: max_bid_ask_spread_pct 0.15 (shared liquidity gate, scanner.py:622-625), documented in config.example.json as "generous enough not to reject normal earnings-week widening but catches genuinely illiquid chains" — a reasoned starting point, not backtested.

Why not act now: one miss on one liquid name doesn't distinguish "threshold too tight" from "this stock's spread genuinely widened that night." Per the protocol below, this needs scan_log data across the accumulation window, not a single anecdote. Loosening it isn't free either — costs.py's slippage haircut scales with spread width, so admitting wider-spread names also directly worsens cost-adjusted expectancy on those trades, partially offsetting any gain from more samples.

Review checkpoint: ~2026-07-23 (two weeks out). Query scan_log (profile='strat_test') for how often bid_ask_spread_too_wide appears as a rejection reason, how close the misses run (median/max overage), and whether it disproportionately hits otherwise-qualifying liquid names (high market cap / high volume) rather than genuinely illiquid ones. If it's a frequent, narrow-miss reason on liquid names, that's the signal to test loosening it (e.g. 0.15 → 0.18); if it's rare or mostly hitting genuinely thin names, leave it as-is.

Prioritized test queue

Evaluate in this order, one parameter at a time, only once a strategy clears the 30-trade directional gate (see docs/strategy-testing-plan.md):

  1. iron_fly vs iron_condor head-to-head expectancy (§2) — no code change, just read the data the paper program already produces.
  2. BWB profit target 0.10 → 0.25/0.50 (§1) — highest-suspicion single value.
  3. Skew logic: validate BWB (§3.2), then prototype skew-leaned iron_condor (§3.1).
  4. IV/RV floor 1.25 → 1.35 (§5) — tightening the core edge.
  5. Calendar front-leg DTE placement (§4).
  6. Stop 1.5× vs 2.0× (§6).
  7. Wing width (§7) — only if needed.
  8. Bid-ask spread gate (§8) — revisit at the 2026-07-23 checkpoint once scan_log shows whether bid_ask_spread_too_wide is a frequent narrow-miss on otherwise-liquid names.

Protocol (do not blind-tune)

  • Change exactly one parameter; keep every other value fixed.
  • Re-run the paper program for a fresh window; compare cost-adjusted expectancy (not gross P&L, not the literature) between the old and new value.
  • Require the change to hold up on ≥30 trades before believing it, ≥100 before trusting it.
  • Generate tearsheets for end-of-window evaluation, never to fine-tune mid-window — that overfits to the specific test window.

Sources

See also

  • docs/strategy-testing-plan.md — the paper program that validates every hypothesis here
  • docs/05-strategies.md — the 7 strategies' structures and current parameters
  • src/strategy_metrics.py — expectancy / IV-crush / regime metrics the tests are judged on