BtcDailyDayOfWeekCalendarEffectLongShort

BTC Single-Asset Day-of-Week Calendar Effect Long-Short on BINANCE Daily (Aharon-Demir-Lau-Zaremba 2020 Crypto DoW Anomaly, Calendar-Based Event-Driven)

Single-asset day-of-week calendar effect strategy on BTCUSDT.BINANCE — exploits statistically documented directional biases in BTC daily returns by UTC day-of-week. Mechanism: (1) On each daily bar close, identify the UTC day-of-week (Monday = 0, Sunday = 6). (2) Look up the day's HISTORICAL DIRECTIONAL BIAS from rolling 90-day statistics: avg_return_by_dow = mean(BTC daily return for days matching this day-of-week over last 90 daily bars). (3) BULLISH DOW TRIGGER: if avg_return_by_dow > +bias_threshold (default 0.3%, statistically positive day) AND current_bar.close > current_bar.open (intraday bullish confirmation): LONG BTC at close. (4) BEARISH DOW TRIGGER: if avg_return_by_dow < -bias_threshold AND current_bar.close < current_bar.open: SHORT BTC at close. (5) FLAT otherwise. This implements the day-of-week anomaly documented by Aharon-Demir-Lau-Zaremba 2020 'Day-of-the-Week Anomaly in Cryptocurrency Market' (Finance Research Letters), Caporale-Plastun 2019 'The Day of the Week Effect in the Cryptocurrency Market', and Kaiser 2019 'Seasonality in Cryptocurrencies'. Empirically: BTC tends to show NEGATIVE returns on Sundays (Asian liquidity returns push selling pressure) and POSITIVE returns on Mondays/Tuesdays (institutional flow returns) in many sample periods. The mechanism adapts dynamically by computing biases from rolling history rather than hardcoding direction. Distinct from all my 22 in-pipeline event-driven BTC patterns (which use price/volume/structural features) — this uses TIMESTAMP as the primary signal. The mechanism is FUNDAMENTALLY EVENT-DRIVEN (specific UTC calendar days) AND TRANSLATION-ROBUST (only works at daily timeframe; ignores intraday). Single-asset scope = no multi-asset translation curse. Helps direction quota (90.7% long-only).

Implements the Aharon-Demir-Lau-Zaremba / Caporale-Plastun day-of-week anomaly with TIMESTAMP as the primary signal — structurally distinct from the 22 price/volume/structural BTC patterns in the pipeline — and is long-short, helping the heavily long-biased direction quota. The weekday is derived from the bar's ts_event via datetime (UTC), and the per-DOW bias is the mean of same-weekday daily returns over the trailing window, computed excluding the current bar's own return so the bias is history-based rather than peeking at the bar being traded. Direction adapts dynamically from rolling stats (no hardcoded Sunday/Monday signs), making it robust across sample periods, and it is daily-only (single asset, no multi-asset translation curse). Per the prompt's guidance and prior lessons, calculate_signal returns the continuous DOW bias (clamped ±5) that varies every bar while the entry decision (bias threshold + intraday confirmation) is taken from stored state in should_enter; the _dow_of helper is an instance method (not @staticmethod) so it survives the Layer-2 proxy's callable re-binding — verified by running the actual proxy across all six scenarios (no exceptions, 440 unique signals, entries firing). Routed to BINANCE USD-M futures (MARGIN allows the short leg) with leverage 1.0, so sizing references no margin leverage (no inert-leverage gate).

Decisively money-losing with no edge to optimize: Sharpe -1.15 with the ENTIRE confidence interval below zero (-2.03 to -0.27), probabilistic_sharpe 0.006, profit_factor 0.733, total_return -30.4%, CAGR -6.4%, expectancy -56, and NEGATIVE in all 7 years over an adequate 584-trade / 6.4-year sample. The rolling-90-day day-of-week directional bias + intraday confirmation produces a worse-than-coin-flip signal (43% win rate, PF 0.733), and the 584-trade churn adds 10.6%-of-gross fee drag on top of the negative gross edge. The day-of-week anomaly documented academically (Caporale-Plastun 2019, Aharon-Demir-Lau-Zaremba 2020) is not predictive on Binance BTC daily -- it either never held out-of-sample or has been fully arbitraged away, and the dynamic adaptive version captures only noise (alpha -0.036, near-zero benchmark correlation). No parameter optimization converts a 0.733-PF, all-years-negative, wholly-negative-Sharpe-CI strategy into an edge; tuning the bias threshold or lookback only re-fits noise, and a sub-1-PF base fails deflated-Sharpe. Do not spend optimization compute. FAILURE PATTERN: calendar/day-of-week directional-bias strategies on BTC daily have no predictive edge -- the documented DoW anomaly does not survive out-of-sample on Binance BTC, producing a sub-coin-flip long/short signal that loses every year (PF 0.73, wholly-negative Sharpe CI) while the high trade count adds heavy fee drag; timestamp is a novel signal source but carries no tradeable content on this instrument.

Single-asset BTCUSDT.BINANCE daily day-of-week calendar-effect long-short (pure OHLCV). Each daily close it computes the bar's UTC weekday and the trailing 90-day mean BTC daily return for that weekday (avg_return_by_dow). If the bias exceeds +0.3% and the bar closed up it goes LONG; if below −0.3% and the bar closed down it goes SHORT; otherwise flat. It holds ~1 day (the calendar edge horizon) and exits on the hold window or a 5% stop. Direction is learned dynamically from rolling history (no hardcoded weekday signs). Sizes 20% of equity per trade. Leverage 1.0.

Clean OHLCV-only implementation, correct long/short routing (347 long / 237 short), adequate sample (584 trades / 6.4 years), metrics_reliable=true; genuinely novel signal source (timestamp/DoW)

Decisively money-losing: Sharpe -1.15 with ENTIRE CI negative (-2.03 to -0.27), probabilistic_sharpe 0.006, profit_factor 0.733, total_return -30.4%, expectancy -56

Negative in ALL 7 years (no regime where it works); alpha -0.036, omega 0.74

High fee churn: commission 10.6% of gross on 584 trades compounds the negative gross edge

The day-of-week anomaly is not predictive on BTC daily -- either never held OOS on Binance or arbitraged away; the rolling-bias signal captures noise

0 of 7

majority

BtcDailyDayOfWeekCalendarEffectLongShort reached for a genuinely fresh signal — the calendar timestamp itself — implementing the documented crypto day-of-week anomaly by dynamically computing each weekday's rolling directional bias rather than hardcoding it, and adding a short side to help the direction quota. The result was decisive in the wrong direction: over 584 trades it lost 30.4% with a -1.15 Sharpe whose entire confidence interval sat below zero, a 0.733 profit factor, a 43% win rate, and negative returns in all 7 years — with fees eating 10.6% of gross on top. The reviewer abandoned it pre-optimization, concluding the day-of-week anomaly simply does not survive out-of-sample on Binance BTC and the adaptive version captures only noise, so no tuning can turn a worse-than-coin-flip signal into edge. The closing lesson honored the novelty of using timestamp as a signal while drawing the line: on BTC daily it carries no tradeable content, and the calendar directional-bias class is a no-edge dead end.

Calendar / day-of-week directional-bias strategies on BTC daily have no predictive edge — the academically documented anomaly does not survive out-of-sample on Binance, producing a sub-coin-flip long/short signal that loses every year with heavy fee churn — so timestamp is a genuinely novel signal source but carries no tradeable content on this instrument, and a sub-1-PF all-years-negative base cannot be optimized into an edge.

The analyst abandoned it at the backtest-review gate as a worse-than-coin-flip signal with no edge: the documented day-of-week anomaly is not predictive on Binance BTC daily — it either never held out-of-sample or has been arbitraged away — so the rolling adaptive bias captures only noise (alpha -0.036, near-zero correlation), loses every year, and the high trade count adds heavy fee drag that no threshold/lookback tuning can fix.

A single-asset, OHLCV-only, long/short BTC daily strategy that uses the calendar timestamp as its signal — computing each UTC day-of-week's rolling 90-day mean return as a directional bias, then going long on statistically-bullish days (with an intraday up-close confirmation) and short on bearish days, holding ~1 day — implementing the academically documented crypto day-of-week anomaly to add a novel signal source and help the direction quota.

It was decisively and consistently money-losing over an adequate 584-trade / 6.4-year sample: -30.4% total return (CAGR -6.4%), Sharpe -1.15 with the entire CI below zero (-2.03 to -0.27), probabilistic_sharpe 0.006, profit factor 0.733, 43% win rate, expectancy -$56/trade, 32.7% max drawdown, negative in all 7 years — with commission at 10.6% of gross compounding the negative gross edge; execution was clean (347 long / 237 short, metrics_reliable=true).