Strategies

LTC Friday-Expiry Weekend Alt-Drift: Long LTCUSDT.BINANCE USD-M From the Close of the 07:00-08:00 UTC Friday Bar for a Fixed 48h, Flat ~71% of the Time (1H bars, long-only, 0.4x notional, pure OHLCV, 1 structural parameter + catastrophe stop)

Outcome: Abandoned

LtcFridayExpiryWeekendDriftLong

Outcome Summary

LtcFridayExpiryWeekendDriftLong went long LTC perps from Friday 08:00 UTC for 48h, based on a pre-study showing a Friday weekend drift across 10 alts. Its first backtest cleared the fee floor comfortably (0.88% per trade, PF 1.22, 345 trades) but missed the pre-registered PF and drawdown targets, and it was still sent to optimization. The optimized parameters produced a near-zero holdout Sharpe (0.03 over 69 trades), a 55% drawdown and a failed FDR test, so the analyst abandoned it as no_edge after one iteration. The analyst also asked for an engine review: the daily MTM series showed large paired swings on days the book was flat by construction, which contaminates this run's Sharpe, drawdown and DSR.

Strategy report

Hypothesis

A single-instrument, single-leg, long-only calendar EVENT trade on the LTCUSDT.BINANCE USD-M perpetual, using 1-HOUR bars and pure OHLCV, with no supplementary feed. Every Friday at the close of the 07:00-08:00 UTC bar (bar timestamp Fri 08:00 UTC, the moment Deribit's weekly BTC/ETH options expire and settle), go long at 40% of equity notional. Exit on a fixed 48h clock at the close of the bar stamped Sunday 08:00 UTC. The position is flat Sunday 08:00 to Friday 08:00, about 71% of calendar time. One trade per week gives ~52 trades/yr and ~350 over the 2020-2026 LTC history. POINT-IN-TIME PRE-STUDY (researcher, Binance 1H closes 2020-01 to 2026-09, next-close execution, 0.10% round trip deducted, no parameters fitted). Friday 08:00 to Sunday 08:00 is the best 48h start day on LTC: mean +0.87% gross, against +0.12% for the other six weekdays' 08:00 starts. Net of fees on LTC: n=351, mean +0.77%/trade, PF 1.51, win 59%, calendar Sharpe 1.01 at full notional. Years: 2020 +113%, 2021 +23%, 2022 -5%, 2023 +46%, 2024 +61%, 2025 +11%, 2026 YTD +22% (sum of trade returns). First 60% of trades average +0.83%, last 40% (from 2024-01) +0.68%, so the effect has not decayed. Clock choice follows L166: the forward return was measured at 24/36/48/72h, and 48h is the peak (on the 10-name basket, 24h Sharpe 0.82, 36h 0.99, 48h 1.12, 72h 0.75). The clock was not chosen to hit a time-in-market target. CROSS-SECTIONAL ROBUSTNESS (why this is not a best-of-N LTC artifact). The same rule was measured on 10 liquid Binance USD-M names: ETH, SOL, XRP, BNB, ADA, LINK, DOGE, DOT, LTC, AVAX. Friday 48h exceeds the other-weekday 48h mean on ALL 10. The pooled Friday excess is positive in EVERY year: 2020 +1.35%, 2021 +0.21%, 2022 +0.16%, 2023 +0.72%, 2024 +0.87%, 2025 +0.27%, 2026 +0.54%. The equal-weight basket is net PF 1.57 with a 48h calendar Sharpe of 1.12. Per-name net Sharpe runs 0.58 (ETH, DOT) to 1.01 (LTC), with a median of about 0.84. LTC is the best of the 10, and that selection is DISCLOSED here: the prediction below uses the cross-name median, not LTC's 1.01. LTC was chosen as the single expression because L173 says to express the mechanism on one liquid instrument rather than a basket that times out, L137/L151 rule out multi-leg books and hand-typed universes, it has among the lowest corpus shares of the liquid majors, it has the most consistent years (6 of 7 positive), and it has full 2020+ Binance history. BTC does NOT carry the effect (BTC Friday 08:00 48h +0.32% vs +0.56% Monday). That is the likely reason the prior 'BTC Spot Weekend Holding Long' failed, and it is what distinguishes this alt-specific version. NOVELTY CHECK (L138, mechanism stripped of instrument): 'hold a liquid ALT long from Friday 08:00 UTC through the weekend for a fixed 48h, flat otherwise'. Prior corpus: BTC spot weekend hold (BTC does not carry it, see above), ETH day-of-week long-short (always-in, both directions, ETH has the weakest excess of the 10), and weekend move-fade reversions (opposite mechanism, entered Monday). No prior run is a long-only, flat-otherwise, alt-perp Friday-anchored 48h clock.

This implements the hypothesis directly. The schedule depends only on the bar's close timestamp: hours since the last Friday 08:00 UTC, computed from the epoch weekday (checked: 2026-09-25 08:00 UTC maps to hour-of-week 104). It has no bar counters, so restarts, backtest and paper all trade the same dates. The entry fires only on the anchor bar (entry_window_hours=1). The exit fires once 48 or more hours have passed since the anchor, which covers a missing Sunday 08:00 bar. calculate_signal returns a continuous weekly phase: 1 at the anchor, 0 at the scheduled exit, negative while flat. Sizing is equity*0.4*leverage/price with leverage=1, so gross is capped at 0.4x. The catastrophe stop is checked at the 1H close in should_exit instead of through a bracket, because a bracket would need an artificial take-profit and the hypothesis has none. I used futures because the hypothesis names the USD-M perp and its perp-fee pre-study. Long-only at 0.4x would also fit BINANCE_SPOT, but at ~0.20% round trip. Expected ~0.77% per trade clears the 0.15% futures floor. The hypothesis names 48h as the peak of the measured holding windows, so this clock is pre-registered and should not be re-tuned by the optimizer. Keep anchor_hour_of_week fixed.

Failure code: no_edge, based on the first measured report plus the holdout. It is not fee_edge: avg_trade_return_pct is 0.88% against the 0.15% floor. The first report showed PF 1.22, Sharpe 0.47 (CI -0.25 to 1.13) and DD 32.3%, which already missed the pre-registered PF of 1.25 and max DD of 25%. After optimization, the unseen holdout (2025-04-21 to 2026-08-16, 69 trades) returned Sharpe 0.03, which is effectively zero forward edge. The optimized book has PF 1.07, DD 55%, impact at 68% of gross, DSR 0.79, and it fails programme FDR. Nine soft floors fail. There is no defensible waiver on a near-zero holdout, and re-optimizing the same calendar mechanism is barred by L126. The mechanism itself is simple (one calendar slot, one clock), so no iterate path exists; only the slot and size can change, and the sweep showed the slot is a cliff (anchor 114 gives Sharpe 0.16). Research Lead note: the Friday-08:00 alt weekend drift did not survive out of sample on the best-selected name (LTC was chosen as the top of 10). Do not re-propose single-name versions, or other anchors or hold lengths. ESCALATE FOR ENGINE REVIEW (suspected artifact, cause not verified): the daily MTM series shows paired -23% to -39% / +30% to +64% swings on days the strategy is flat by construction (e.g. baseline 2022-01-25/26, 2023-08-22/23, 2025-10-21/22). There is also a -37.96% print on the flat final bar (2026-08-17), and the optimized series has a -84%/+535% pair and ends at -100% on 2025-05-25 at the holdout boundary. These contaminate Sharpe, drawdown and DSR for this run, and possibly others on LTCUSDT.BINANCE 1H.

Implementation

Calendar event trade on LTCUSDT.BINANCE USD-M perp, 1H bars, pure OHLCV. Goes long 40% of equity notional at the close of the bar stamped Friday 08:00 UTC (the 07:00-08:00 bar, Deribit weekly expiry) and exits on a fixed 48h clock at the close of the bar stamped Sunday 08:00 UTC. Flat the rest of the week (~71% of calendar time). A 15% close-checked catastrophe stop is the only other exit. About 52 trades a year.

Verification Results

Keep the pre-registered rule as it is. The analyst should read the worst-trade loss and drawdown from the full-history backtest.

Per-trade risk (L148). At 0.4x notional with a 15% catastrophe stop checked at the close, the nominal worst loss is about 6% of equity per trade, and more on a gap. That is above the 1-2% guideline. Mitigating factors: the design is not always-in (flat about 71% of the time), it has a fixed 0.4x gross cap, no dynamic vol scalar, and leverage 1.0. Worst-case gross = 1 leg x 0.4 x 1.0 = 0.4x equity, well under the venue cap.

No change needed.

The entry fires only on the bar stamped exactly Fri 08:00 UTC (hours_since_anchor 0 < entry_window_hours 1). If that single 1H bar is missing from the data, the week is skipped, and the code does not enter late. That behavior is correct and point-in-time safe; it is noted only so that a slightly lower trade count is not mistaken for a bug. The exit uses >= 48, so a missing Sunday 08:00 bar exits on the next bar.

Exclude anchor_hour_of_week, hold_hours and entry_window_hours from the optimization search space, or treat them as fixed.

anchor_hour_of_week and hold_hours are exposed in self.parameters. Per the hypothesis and the developer rationale, they are pre-registered structural constants (Fri 08:00, 48h). If the optimizer searches them, it would turn a disclosed single-rule test into a best-of-N calendar search.

The sandbox (834 days, 119 trades) shows avg_trade_return_pct of 0.68%, which is close to the pre-study's +0.77%. PF is 1.24, but the Sharpe CI (-1.15 to 1.45) straddles 0. The calm-vol tercile lost money (-7.3%). Impact takes 33.8% of gross, which is high for a 0.4x LTC book, so check capacity and impact at full size. Funding is NOT folded (funding_events_available=0). A long held over the weekend pays about 6 funding stamps per 48h hold, and in bull regimes that can cost 0.03-0.1% or more per trade. The pre-study also omitted funding, so the full backtest should confirm the per-trade net after funding. The LTC choice is a disclosed best-of-10 pick. Judge against the cross-name median Sharpe of about 0.84, not LTC's 1.01. The single skew outlier is the largest loss, -$11.75k, about 3x the largest win.

Backtest Review

Sharpe: 0.47
Total return: 96.99%
Max drawdown: 32.33%
Trades: 345
Win rate: 57.1%
Profit factor: 1.22

The trades match the hypothesis. All 345 are long, entered at the close of the Friday 08:00 UTC bar and held 48h or until the catastrophe stop. Every signal was submitted.

Per-trade edge clears the fee floor. The ledger's mean net pnl_pct is 0.49% and the report's avg_trade_return_pct is 0.88%, against a 0.15% floor, so this is not fee_edge. PF is 1.22 and the win rate 57%.

The trade ledger adds up: the summed pnl ($196,995) equals the final equity. end_unrealized_pct is 0.

The return is not market beta (L170): beta 0.10, exposure 28%, and a 32% drawdown against roughly 80% or more for LTC buy-and-hold. Alpha is +8.6% a year.

Funding is now included and costs $20.5k in total. The per-trade edge survives it.

Suspected engine artifact (mechanism not verified). daily_returns shows paired swings of about -23% to -39% followed by +30% to +64% on days the book is flat (for example 2023-08-22/23, 2025-10-21/22, 2026-06-24/25). The final flat bar, 2026-08-17, shows -37.96% and never reverses. Sharpe 0.47, its confidence interval, PSR, the regime table and annual_returns (2026 shows -36.8%; the ledger has about +4%) are all built on this series and cannot be trusted.

The full backtest is weaker than the pre-study. Ledger results by year: 2020 +94%, 2021 +0.5%, 2022 -22%, 2023 +40%, 2024 +51%, 2025 +1.7%, 2026 +4%.

The average loss is larger than the average win. The largest single loss is -24%, when price gapped through the 15% stop on 2025-10-10.

Market impact takes 35.9% of gross profit. Capacity is about $0.78M.

Realized leverage runs 0.40-0.56 against the pre-registered 0.4x.

The book is flat about 72% of the time. L165 doubts such a book can reach a Sharpe of 1.0; that is for the promotion stage.

Analysis

Benjamini-Hochberg at q=0.10 over 339 programme candidates keeps 8. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

A calendar effect found on the best of 10 names (LTC) did not hold out of sample, so the Research Lead should not re-propose single-name versions or variants with a different anchor or hold length; separately, the flat-day daily-MTM swings seen in this run need an engine review before LTCUSDT.BINANCE 1H metrics can be trusted.

Backtest review sent it to optimization, but after optimization the analyst abandoned it with code no_edge: the holdout Sharpe of 0.03 showed essentially no forward edge, nine soft floors failed, and the initial report had already missed the pre-registered PF of 1.25 and max drawdown of 25%. Re-optimizing was ruled out by L126, and the sensitivity sweep showed the anchor slot is a cliff (anchor 114 gave Sharpe 0.16).

A long-only calendar event trade on the LTCUSDT.BINANCE USD-M perp on 1H bars: buy at 0.4x equity notional at the close of the Friday 08:00 UTC bar (weekly Deribit expiry), hold for a fixed 48h, and stay flat the rest of the week, betting on a weekend drift in alts that the pre-study found on all 10 liquid names tested.

The initial backtest (2020-01 to 2026-08, 345 long trades) returned +97.0% with Sharpe 0.47 (CI -0.25 to 1.13), PF 1.22, 57% win rate, 32.3% max drawdown and avg_trade_return_pct 0.88%. The optimized version (anchor 102, hold ~52h, ~0.58x notional) had WF OOS Sharpe 0.66 and PBO 0.15, but its unseen holdout (2025-04-21 to 2026-08-16, 69 trades) returned Sharpe 0.03, with PF 1.07, 55% drawdown, impact at 68% of gross, and it failed programme FDR (p=0.21).

Backtest and paper results are hypothetical. Trading involves risk of loss.