WeeklyThursdaySeasonalityShortFundingFiltered

Weekly Thursday Seasonality Short with Funding Filter

Exploit weekly seasonality pattern where crypto tends to sell off on Thursdays, likely driven by Deribit options expiry hedging (Friday 08:00 UTC), weekly funding rate cycles, and institutional end-of-week risk reduction. Entry: Open short on Thursday (around 00:00-12:00 UTC) on Binance USD-M Futures, but ONLY when the current funding rate is positive (indicating crowded longs and liquidation potential). Exit: Close position Friday morning before options expiry (~08:00 UTC), or after a fixed 12-24h hold period. Stop-loss at 1.5-2% above entry to cap upside risk. Instruments: BTCUSDT and ETHUSDT on Binance Futures (MARGIN account). Can expand to other large-cap alts if the pattern holds. Supplementary data needed: funding_rates for entry filtering. Optionally open_interest to confirm elevated positioning. Expected edge: 0.2-0.5% avg per-trade return after fees (~0.10% round-trip). Low frequency (~1 trade/week per instrument) keeps fee drag minimal. Risk factors: Seasonality may be regime-dependent and decay over time. Thursday rallies during strong bull trends could trigger stops. Small sample size per year (~52 trades) requires multi-year validation.

Iteration-4 fix (Layer 3 — sandbox smoke test exceeded 300s wall-clock). The funding lookup was already O(log n) via a precomputed sorted array + searchsorted, so the timeout came from unconditional per-bar overhead: _manage_all() ran on every primary AND extra bar and, for every asset on every bar, called get_account_equity(), cache.positions_open(), built a list, and constructed a tz-aware datetime — even on the ~99% of idle bars. Two smallest-possible changes eliminate that: (1) _wd_hour now derives weekday/hour with pure integer arithmetic (no datetime allocation, using the fact that 1970-01-01 was a Thursday=index 3); (2) a perf gate in _manage_all skips ALL cache/equity work unless the bar is inside the Thursday entry window or the asset is currently held (tracked in self._active, set on entry and cleared on exit). Idle bars now cost only a couple of integer ops, independent of history length. All signal/entry/exit logic, funding gating, sizing, caps and imports are unchanged, so Layers 1–2 remain green; correctness is preserved because positions only ever exist via our own entry path (which always adds to self._active), and held assets still run full exit checks every bar.

failed deflated Sharpe: DSR=0.0, expected-max=3.9052 over 225 trials — the selected optimized Sharpe (0.801) sits FAR below the best-of-N luck bar and is statistically indistinguishable from noise (is_significant=false, sharpe_ci_low=-0.453 straddling 0; probabilistic_sharpe 0.9191 vs DSR 0.0 is the textbook PSR-vs-DSR selection-inflation trap). Decisively, sensitivity FAILED with 2 CALENDAR cliffs (entry_weekday, exit_weekday): the heatmap shows Sharpe of 0.65 exists ONLY at entry=Thursday(3)/exit=Friday(4) and inverts to -1.71/-3.81 (entry) and -2.00/-3.81 (exit) one day either side — the classic single-cell seasonality curve-fit, i.e. the Thursday-selloff premise is a fitted day-of-week coordinate, not a robust pattern. The walk-forward is is_overfitted=TRUE: avg IS 0.95 collapses to avg OOS -1.95 with OOS windows [0.83, -7.02, 0.34], a catastrophic middle fold. The holdout 'pass' is SPURIOUS — ratio 0 measured against a NEGATIVE WF-OOS baseline (-1.95), so holdout_sharpe 1.86 is not real forward validation. benchmark_meaningful=false, so on absolute metrics it still fails outright: Sharpe 0.65/0.80 << 1.5, max_dd 36%/24.7%, negative information ratio (-0.48/-0.54), max_drawdown_duration ~1085-1274 days. Not iterate (attempt 1 of 2): the only positive-Sharpe region IS the frozen calendar coordinate that constitutes the overfit, so there is no robust parameter region to tune toward and a second best-of-225 sweep re-selects the same single-cell fit. Not revise_hypothesis: per the overfit->abandon rule this is high-IS/negative-OOS overfitting on valid instruments (BTC/ETH), with no promoted sibling stranded on a dead target — re-pointing a day-of-week seasonality onto another day or asset just re-fits the calendar. FAILURE PATTERN: a day-of-week 'Thursday selloff' short gated by positive funding on BTC/ETH does NOT produce a promotable edge — the sensitivity surface exposes it as a single-cell calendar fit (Sharpe positive ONLY at Thu-in/Fri-out, strongly negative one day either side), and it fails every forward gate (DSR 0.0 with optimized Sharpe 0.80 below a 3.91 luck bar, is_overfitted with a -7.02 OOS window, a holdout 'pass' inflated by a negative WF-OOS baseline). A low market beta and a plausible microstructure narrative measure diversification and story, not significance or forward-generalization.

Shorts the weekly Thursday seasonality on BTCUSDT and ETHUSDT (Binance USD-M, 1H bars), gated by positive funding (crowded longs). Enters SHORT during the Thursday 00:00-12:00 UTC window only when that asset's funding rate is positive; covers Friday 08:00 UTC before Deribit expiry, on a max-hold time-stop, or on a fixed-% stop above entry. Stop-distance-based sizing (risk_frac of equity), per-position notional cap, and a 1.0x gross cap (no leverage).

Verification failed (Layer 3 — sandbox backtest): smoke test exceeded the 300s wall-clock limit. This almost always means per-bar work that scales with history — e.g. rescanning the full funding/supplementary series, or rebuilding a list and calling min()/sorted() inside calculate_signal()/on_bar() on every bar. Precompute sorted timestamp arrays ONCE in __init__ and use bisect, or cache lookups keyed by timestamp, so per-bar cost is O(log n) not O(n).

Large, clean sample: 722 short trades over 2020-2026, no liquidation, full history — trades match the hypothesis (all shorts, Thursday entry ~18h hold, funding-gated)

Positive net-of-fees edge distributed across years: total_return 86.4%, CAGR 11.1%, PF 1.19, expectancy +$136/trade, positive in 5 of 7 years

Low/negative beta (-0.15) and benchmark correlation (-0.37) — a genuine short-seasonality diversifier vs the long-biased pipeline

3 clean tunables with real room to improve selectivity (funding_min, stop_pct, max_hold_hours)

Thin edge: Sharpe 0.654 with sharpe_ci_low -0.65 (CI straddles 0), PF 1.19, information_ratio -0.48 — may not survive deflated-Sharpe/walk-forward

Heavy cost drag: impact_cost_pct 17.8% of gross PnL and capacity_usd only $3.14M, commission 7.78% of gross — edge exists only at modest size

Seasonality decay risk: 2024 negative (-8.8%), and the hypothesis itself flags regime dependence; OOS generalization is the real question

High kurtosis (77) / skew 5.7 — some reliance on outlier short days (e.g. 2020-03-13 COVID crash +42.6%)

Genuinely low/negative market beta (base -0.15, opt -0.11) and negative benchmark correlation (-0.37/-0.45), so the book is directionally short-biased as intended, not a closet long.

Fee drag is modest (commission_pct_of_gross ~6.9-7.8%) and per-trade edge clears costs on paper; funding filter and gross cap are implemented cleanly.

Sensitivity FAILED with 2 calendar cliffs (entry_weekday, exit_weekday): Sharpe of 0.65 exists ONLY at Thursday-in/Friday-out; one day either side collapses to -1.7 to -3.8 — the classic single-cell seasonality overfit signature.

Deflated Sharpe = 0.0 (vs 0.95); optimized Sharpe 0.801 far below the 225-trial expected-max luck bar of 3.905; is_significant=false; sharpe_ci_low=-0.453 straddles 0.

Walk-forward is_overfitted=TRUE: avg IS 0.95 collapses to avg OOS -1.95 with windows [0.83, -7.02, 0.34] — a catastrophic middle fold.

Holdout 'pass' is spurious: ratio 0 measured against a NEGATIVE WF-OOS baseline (-1.95); not real forward validation.

Absolute risk-adjusted metrics miss promotion: Sharpe 0.65/0.80 << 1.5, max_dd 36%/24.7%, negative information ratio (-0.48/-0.54), max_drawdown_duration ~1085-1274 days.

This strategy shorted a claimed Thursday sell-off in BTC/ETH, gated by positive funding, and unlike most siblings it survived the backtest-review gate ('optimize') on the strength of an 86.4% total return, PF 1.19, and a genuinely negative market beta. Optimization then exposed it: the positive Sharpe existed only at the frozen Thursday-in/Friday-out coordinate and inverted sharply one day either way, deflated Sharpe was 0.0 with the optimized Sharpe (0.80) far below a 3.91 luck bar, and walk-forward flipped from +0.95 in-sample to -1.95 out-of-sample. The analyst abandoned it after 4 iterations as a textbook single-cell calendar overfit with no robust parameter region to tune toward and no promoted sibling stranded on a dead target. It reached the ANALYZING stage but was never promoted to paper.

A day-of-week seasonality that is profitable only at one exact entry/exit calendar cell is a single-cell curve-fit, not a robust edge — low beta and a plausible microstructure story measure diversification and narrative, not statistical significance or out-of-sample generalization.

The analyst's post-optimization verdict was 'abandon': the sensitivity surface showed 2 calendar cliffs (positive Sharpe existed ONLY at Thursday-in/Friday-out, collapsing to -1.7 to -3.8 one day either side), deflated Sharpe was 0.0 against a 3.91 best-of-225 luck bar, and walk-forward was overfitted (avg IS 0.95 → avg OOS -1.95 with a -7.02 fold).

A low-frequency short-seasonality trade on Binance BTC/ETH that opened a short on Thursday (00:00–12:00 UTC) only when funding was positive (crowded longs) and covered Friday morning before Deribit expiry, betting on a recurring end-of-week sell-off.

The initial backtest looked promising over 722 short trades — total return 86.4%, CAGR 11.1%, profit factor 1.19, expectancy +$136/trade, and a diversifying beta of -0.15 — but the edge was thin (Sharpe 0.654 with CI straddling zero), and after optimization it held Sharpe just 0.801 while failing every forward-validation gate.

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['assets', 'stop_pct', 'exit_hour', 'risk_frac', 'funding_min', 'exit_weekday', 'min_notional', 'entry_weekday', 'qty_precision', 'entry_end_hour', 'gross_cap_frac', 'max_hold_hours', 'entry_start_hour', 'max_position_frac', 'default_qty_precision'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: super(type, obj): obj must be an instance or subtype of type (bar timestamp: 1735689600000) - steady_downtrend: TypeError: super(type, obj): obj must be an instance or subtype of type (bar timestamp: 1735689600000) - flat_ranging: TypeError: super(type, obj): obj must be an instance or subtype of type (bar timestamp: 1735689600000) - volatility_spike: TypeError: super(type, obj): obj must be an instance or subtype of type (bar timestamp: 1735689600000) - zero_volume: TypeError: super(type, obj): obj must be an instance or subtype of type (bar timestamp: 1735689600000) - price_gap: TypeError: super(type, obj): obj must be an instance or subtype of type (bar timestamp: 1735689600000)

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['assets', 'stop_pct', 'exit_hour', 'risk_frac', 'funding_min', 'exit_weekday', 'min_notional', 'entry_weekday', 'qty_precision', 'entry_end_hour', 'gross_cap_frac', 'max_hold_hours', 'entry_start_hour', 'max_position_frac', 'default_qty_precision'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: WeeklyThursdaySeasonalityShortFundingFiltered._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735691340000) - steady_downtrend: TypeError: WeeklyThursdaySeasonalityShortFundingFiltered._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735691340000) - flat_ranging: TypeError: WeeklyThursdaySeasonalityShortFundingFiltered._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735691340000) - volatility_spike: TypeError: WeeklyThursdaySeasonalityShortFundingFiltered._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735691340000) - zero_volume: TypeError: WeeklyThursdaySeasonalityShortFundingFiltered._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735691340000) - price_gap: TypeError: WeeklyThursdaySeasonalityShortFundingFiltered._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735691340000)