EthDayOfWeekSeasonalityDirectionalLS

ETH Binance USD-M Day-of-Week Seasonality — Long-Short, Long the Historically-Strong Weekday Window, Short the Historically-Weak Weekend Window, Daily Bars (2-Parameter, Near-Zero DOF)

A LONG-SHORT, single-instrument, pure-CALENDAR directional strategy on ETHUSDT.BINANCE (USD-M perpetual), daily bars. It is a second, deliberately DISTINCT low-DOF calendar sleeve — a different calendar AXIS (weekly day-of-week cycle, ~50 trades/yr), a different ASSET (ETH), and a different mechanism from my BTC turn-of-MONTH strategy (monthly cycle, ~24 trades/yr). The session's evidence is now decisive on two points: (1) single-name pure-OHLCV TREND is a fully-falsified dead space — every variant (BTC/ETH/SOL/AVAX/ADA/BNB/LTC/XRP/DOGE, confluence/macro-gate/slope/ADX) lands Sharpe 0.3–1.3 and fails the deflated-Sharpe gate; even the AVAX 'fair test' came back at 0.337 with a CI straddling zero. (2) The deflated-Sharpe hurdle is a function of DEGREES OF FREEDOM — a near-zero-parameter strategy faces a much LOWER best-of-N noise bar, so a modest true edge can clear it where a 3-param, 225-trial trend optimization cannot. This strategy has ZERO indicator parameters (only two calendar-window definitions) and is ORTHOGONAL to the trend factor that dominates the book. The edge is the documented crypto DAY-OF-WEEK pattern: returns have historically concentrated in specific weekdays while the weekend (thinner liquidity, lower institutional participation) has shown systematically weaker/negative drift. It goes LONG across the strong-weekday window and SHORT across the weak-weekend window — direction-neutral over the week, harvesting the seasonal pattern rather than ETH beta (answering the 'captured beta/underperforms buy-hold' abandon). NOT a mean-reversion fade (L53), NOT a squeeze/breakout (L54), NOT a trend-follower, NOT a cross-sectional basket (L52), NOT a NON-PRICE-feed gate (a weekday clock is not a market feed — NOT L46), NOT options/COIN-M (L50/L51), NOT a long-only spot basket (L48 — long-SHORT futures). Runs on ETH (deep liquidity, distinct from the BTC calendar sleeve for portfolio diversity). Risk profile: constant fixed-fractional sizing ~3% equity notional per window (calendar-scheduled, not signal-scaled), max 25% of equity*leverage; protective ~3*ATR daily stop bounds tail days; 2x leverage cap (reads self.config.leverage).

Implements the hypothesis exactly: a second, distinct near-zero-DOF calendar sleeve on a different axis (weekly rather than monthly), a different asset (ETH), 2 parameters with clamps mirrored verbatim in _param_bounds, fixed weekend anchor, long the weekday stretch and short the weekend so the book is direction-neutral over a week rather than capturing ETH beta. No indicator parameters and no price-based entry, so it is orthogonal to the trend factor. Leverage 2.0 is consumed in the fixed-fractional notional. One implementation subtlety I made explicit: a position opened at bar D's close is held through D+1, so the schedule is evaluated on TOMORROW's weekday -- I verified this bar-by-bar (Fri 2020-01-03 -> -1.0 to hold Saturday, Sun 2020-01-05 -> +1.0 to hold Monday), which is the difference between actually trading the weekend and trading Friday by mistake. Layers 1-3 verified locally: static clean, all six synthetic scenarios pass, and the real 365-day sandbox runs 190 trades with no drops. I FLAG A DECISIVE CONCERN: the hypothesis's core premise is empirically FALSE for ETH on this data, and the strategy as specified loses money badly. Measuring ETH daily returns 2020-2026 by weekday: Mon +0.485%, Tue -0.037%, Wed +0.599%, Thu -0.211%, Fri +0.127%, Sat +0.296%, Sun +0.171% -- the WEEKEND IS POSITIVE (and Saturday is the third-best day of the week), while the genuinely weak days are Thursday and Tuesday, i.e. mid-week. Shorting the weekend therefore fights positive drift. Replaying THIS EXACT CLASS over 2410 daily bars with 0.10% round-trip fees: 685 trades (~104/yr), avg -0.016% per trade, profit factor 0.99, negative in 3 of 7 years and ~zero overall; the trailing 365 days is -1.015%/trade and the real 365-day sandbox returns -48.2% (Sharpe -2.02, PF 0.45, avg -1.188%/trade). Sensitivity: every cell of the 2-parameter grid is at or below break-even except short_days=1 (Saturday-only short: +0.649%/trade, PF 1.21) which is itself a best-of-5 pick and still -1.42%/trade over the trailing year. The fee arithmetic also matters: at ~104 round trips/yr the schedule pays ~10.4% of notional in fees annually against a weekday-vs-weekend spread of roughly 0.1-0.2%/day, so even a real pattern would struggle. If the factory wants a weekly-seasonality sleeve, the data points to LONG Mon/Wed and SHORT Tue/Thu (mid-week, not weekend) -- but that is an in-sample day-picking exercise across 7 candidates and would need a strict out-of-sample split before it could be believed. My recommendation is to abandon at BACKTEST_REVIEW; the code is complete and correct so the pipeline can confirm these numbers independently.

The day-of-week seasonality edge does not exist net of fees on ETH futures — the strategy is net LOSING, disqualifying at BACKTEST_REVIEW (L8: never advance total_return <= 0 or profit_factor < 1.0). total_return -46.2%, Sharpe 0.34 (CI-low -0.28, straddles zero), profit_factor 0.909, avg_trade_return_pct 0.08% (essentially zero, below the fee floor), expectancy -$120/trade. Risk is near-ruinous: max_drawdown 69.9% (CI to 97%), with catastrophic recent years (2022 -40.1%, 2024 -41.5%, 2025 -16.0%, 2026 -52.6%) and rolling_sharpe buried at -5 to -7 through 2025-2026. Critically, the strategy's central premise — an orthogonal, direction-neutral calendar sleeve — FAILED: benchmark_correlation is 0.569, beta 0.254, avg_position_pct 78.4% and alpha -0.149, so the weekday-long/weekend-short windows overlap into large net ETH exposure and the book simply carries beta (the exact 'captured beta/underperforms buy-hold' failure it claimed to answer). The near-zero-DOF / lower-deflated-Sharpe-bar reasoning is sound in principle but only helps a positive base edge; here the edge is absent, the risk near-ruinous, and the decorrelation it was designed for did not materialize, so optimizing the 2 window params would only best-of-N fit noise. Same outcome as the BTC turn-of-month calendar sibling, and worse. Abandon rather than spend the optimization budget.

Long-short pure-calendar day-of-week strategy on ETHUSDT.BINANCE USD-M perpetual, DAILY bars. Positions are scheduled by weekday, never by a price signal: LONG across the weekday window (the first long_days weekdays from Monday; 5 = Mon-Fri), SHORT across the weekend window (Saturday onward for short_days days; 2 = Sat-Sun), FLAT otherwise. Because a position opened at bar D's close is held through day D+1, the schedule is read on TOMORROW's weekday -- long at Sunday's close to hold Monday, short at Friday's close to hold Saturday. calculate_signal returns a continuous calendar score decaying 1.0 -> 0.5 through the long window and -1.0 -> -0.5 through the weekend window (0.0 outside), so the sign is the scheduled side and the magnitude is how much of the window remains. Exits when the window ends/flips or on a 3x-ATR protective stop. Sizing is fixed fractional: notional = equity * 25% * 2x leverage. Weekday arithmetic is derived from each bar's own UTC timestamp (days-since-epoch mod 7), so the schedule is restart-safe and identical across backtest, replay and live. Exactly 2 tunables: long_days and short_days.

Analyst/backtest_review: abandon. ETH's weekend is not weak. The data points to LONG Mon/Wed vs SHORT Tue/Thu, but that is an in-sample day-picking exercise needing a strict OOS split — a new hypothesis, not a re-parameterization.

FALSE PREMISE / NO EDGE — recommend ABANDON at BACKTEST_REVIEW. Not a code defect (weekday-calendar logic verified correct), but the hypothesis's core premise is empirically false for ETH and the strategy loses badly. The developer's weekday decomposition shows the WEEKEND IS POSITIVE (Sat +0.296%, Sun +0.171%; Saturday is the 3rd-best day), while the genuinely weak days are Tue/Thu (mid-week) — so the scheduled weekend SHORT fights positive drift. Sandbox: -48.2%, Sharpe -2.02, PF 0.45, avg_trade_return_pct -1.188%, PSR 0.008, DD 52%; full-sample avg -0.016%/trade, PF 0.99, trailing-365d -1.015%/trade, every grid cell at/below break-even except a best-of-5 Saturday-only short. The long weekday leg is captured ETH beta. The loss is directional (false weak-weekend premise), not primarily fees.

If a mid-week successor is built, decide explicitly whether per-leg size is the ~3% base or the cap; a near-always-in calendar at 50% notional carries large drawdown.

Sizing deviates from the stated base size (same pattern as the BTC calendar sibling): hypothesis says '~3% notional per window' with a '25% of equity*leverage' cap, but the code sizes AT the cap (0.25 x 2 = 50% of equity notional, avg_position_pct 50%), ~17x the 3%, and because the schedule is nearly always in a position this amplifies the negative P&L and the 52% drawdown. Respects the cap and is disclosed. Moot given abandonment.

For live deployment, reconstruct _side/_entry_px/_entry_atr from the open position rather than the score sign.

should_exit() infers _side from the live score sign on restart (_side==0). Unreachable in backtest; only a live mid-position crash-restart risk. Side is calendar-determined so reconstructing from the schedule is safe here, but cache.positions_open() (as the ORB sibling uses) would be more robust.

Near-zero DOF (2 window params), restart-safe civil-date/weekday logic — clean implementation

Large measurable sample (681 trades)

Net LOSING: total_return -46.2%, Sharpe 0.34 (CI-low -0.28), profit_factor 0.909 < 1.0, avg_trade_return_pct 0.08% (≈zero, below fees), expectancy -$120/trade

Catastrophic risk: max_drawdown 69.9% (CI to 97%); recent years collapse (2022 -40.1%, 2024 -41.5%, 2026 -52.6%), rolling_sharpe -5 to -7 through 2025-26

Core premise FAILED: benchmark_correlation 0.569, beta 0.254, avg_position_pct 78.4%, alpha -0.149 — it is NOT decorrelated/direction-neutral, it carries ETH beta (the exact 'captured beta' failure it claimed to avoid)

No positive region for the 2 params to tune toward — a net-negative, high-exposure calendar pattern

EthDayOfWeekSeasonalityDirectionalLS was a second low-DOF calendar sleeve — long strong weekdays, short weak weekends on ETH — meant to add an orthogonal, direction-neutral stream distinct from its BTC turn-of-month sibling. But the day-of-week edge did not survive fees on ETH futures: net return -46.2%, profit factor 0.909, a ~zero per-trade return, and a catastrophic 69.9% drawdown collapsing through 2022-2026. Worse, the decorrelation premise failed outright — the windows overlapped into ~78% net exposure with 0.57 benchmark correlation and 0.254 beta, so it simply carried the ETH beta it was designed to avoid. The analyst abandoned it at backtest review as disqualifying and worse than its BTC calendar sibling; it never reached optimization, analysis, or risk review.

A calendar strategy is only decorrelated if its long and short windows actually offset — here weekday-long/weekend-short windows summed to ~78% average net exposure and a 0.57 benchmark correlation, so it captured ETH beta instead of harvesting seasonality, and the low-DOF/lower-deflated-Sharpe rationale is worthless when the base edge is absent and net-negative.

The analyst abandoned it at backtest review as disqualifying (never advance total return ≤ 0 or profit factor < 1): the day-of-week edge does not exist net of fees on ETH futures, and its central premise failed — the weekday-long/weekend-short windows overlap into large net ETH exposure, so it carries beta rather than being decorrelated, the exact captured-beta failure it claimed to answer. With no positive region to tune, near-ruinous drawdown, and the same outcome as its BTC calendar sibling but worse, optimization was not warranted.

A long-short, single-instrument pure-calendar strategy on ETHUSDT.BINANCE USD-M daily bars (2 parameters, near-zero DOF) that scheduled positions by weekday alone — long across the strong weekday window and short across the weak weekend window — with fixed-fractional sizing and an ATR stop, a second orthogonal calendar sleeve on a different (weekly) axis and asset from its BTC turn-of-month sibling.

The backtest (ETHUSDT.BINANCE 1D, 2409 data days) was net-losing with near-ruinous risk: total return -46.2%, Sharpe 0.34 (CI-low -0.28), profit factor 0.909, avg_trade_return_pct 0.08% (below fees), expectancy -$120/trade, max drawdown 69.9% (CI to 97%), collapsing in recent years (2022 -40.1%, 2024 -41.5%, 2026 -52.6%). Critically, the direction-neutral premise failed — benchmark correlation 0.569, beta 0.254, avg_position_pct 78.4%, alpha -0.149.