Crypto Top-Tercile LONG-ONLY – Backtest

Hypothesis

This hypothesis is the FAITHFUL re-implementation that the analyst's post-mortem on the abandoned `Crypto Cross-Sectional Momentum Top-Bottom Tercile` strategy explicitly requested. That strategy failed catastrophically (-98% in 2021 alt-season) because the SHORT LEG with equal-notional sizing got annihilated when DOGE/XRP/ADA went parabolic. The analyst's recommendations were precise and actionable: (a) 'Consider LONG-ONLY momentum instead of long-short. The short leg in crypto momentum has consistently……Show moreShow less

This hypothesis is the FAITHFUL re-implementation that the analyst's post-mortem on the abandoned `Crypto Cross-Sectional Momentum Top-Bottom Tercile` strategy explicitly requested. That strategy failed catastrophically (-98% in 2021 alt-season) because the SHORT LEG with equal-notional sizing got annihilated when DOGE/XRP/ADA went parabolic. The analyst's recommendations were precise and actionable: (a) 'Consider LONG-ONLY momentum instead of long-short. The short leg in crypto momentum has consistently destroyed long-leg gains across multiple regimes; long-only avoids this entirely while still capturing the documented effect.' (b) 'Use volatility-weighted (risk-parity) sizing' per Liu-Tsyvinski 2021 and Bianchi-Dickerson 2023. (c) 'Add per-position stop-loss at -15% (equity-protective, not signal-altering).' (d) 'Use a wider universe (20-30 symbols, top-30 by market cap) to reduce concentration risk.' This proposal implements ALL FOUR recommendations: 15-symbol universe (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, AVAX, LINK, DOT, MATIC, TRX, LTC, UNI, ATOM — top-15 by market-cap with multi-year data), LONG-ONLY (no short leg, no alt-season blowup risk), volatility-weighted positions (each name contributes equal portfolio variance via inverse-vol sizing), and -15% stop-loss per position. 90-day formation (vs 60-day in the failed version) better matches the documented 'sustained narrative cycle' time horizon for crypto. Monthly rebalance. Fills the MACRO 1w+ horizon quota (0.5% — by far the most under-represented bucket) and the multi-instrument scope (8.1%). Long-only is over-represented overall, but the long-only flavor of cross-sectional momentum is academically the ONLY version that actually works in crypto per the abandoned strategy's post-mortem — this is justified, not gap-filling for its own sake.

Iteration 2 makes exactly the one change QA required and nothing else. The monthly cadence no longer uses the process-relative bar counter with the huge-negative sentinel (checklist #18): that made the first bar after any restart fire an immediate full 15-name rebalance — expensive in a crash-restart loop — and anchored the rotation grid to process uptime so it drifted off any calendar grid and could not be reproduced in walk-forward or paper replay. The grid is now derived purely from the bar's own timestamp……Show moreShow less

Iteration 2 makes exactly the one change QA required and nothing else. The monthly cadence no longer uses the process-relative bar counter with the huge-negative sentinel (checklist #18): that made the first bar after any restart fire an immediate full 15-name rebalance — expensive in a crash-restart loop — and anchored the rotation grid to process uptime so it drifted off any calendar grid and could not be reproduced in walk-forward or paper replay. The grid is now derived purely from the bar's own timestamp (`day_index = ts // 86_400_000_000_000`; rebalance when `day_index % rebalance_days == 0` on a day not yet rebalanced), giving identical rotation dates across restarts, windows, holdout and paper. Everything else — the aligned-timestamp ranking, inverse-vol risk-parity weights, gross exposure, the every-bar -15% stop, the universe (with ETC standing in for delisted MATIC) and all parameters — is byte-identical in behaviour. The beta-vs-alpha and drawdown concerns QA raised were explicitly marked as full-backtest questions for the analyst, not a QA block, so I deliberately did not alter the mechanism to chase them: judging this strategy's alpha against the equal-weight basket benchmark is the analyst's call on real data.

The long-only redesign correctly removes the short-leg alt-season blowup, but what remains is levered alt-cycle beta with a promotion-disqualifying drawdown — not worth 2 hours of optimization. max_drawdown is 70.2% (CI to 91.4%), well past the L19 50% hard line, with stress windows of -48% (rate-shock 2022), -30% (Luna) and -33% (FTX); it will not clear the drawdown floor after tuning. It shows no genuine edge over its own basket: alpha 0.08, beta 0.47, benchmark_correlation 0.68, information_ratio -0.44, and……Show moreShow less

The long-only redesign correctly removes the short-leg alt-season blowup, but what remains is levered alt-cycle beta with a promotion-disqualifying drawdown — not worth 2 hours of optimization. max_drawdown is 70.2% (CI to 91.4%), well past the L19 50% hard line, with stress windows of -48% (rate-shock 2022), -30% (Luna) and -33% (FTX); it will not clear the drawdown floor after tuning. It shows no genuine edge over its own basket: alpha 0.08, beta 0.47, benchmark_correlation 0.68, information_ratio -0.44, and Sharpe 0.655 with bootstrap CI [-0.144, 1.412] straddling zero — the 401% headline is the 2023/2024 alt-cycle (2022 -65%, 2025 -32%). The headline is further inflated by mark-to-market and an end artifact: end_unrealized_pct 193.9 means roughly half of total_return is open-position unrealized, and the final day 2026-09-02 alone prints +45.5%; avg_trade_return_pct 13.2% and exposure_pct 436% are the L36 leverage/compounding signature. This is the L45/L49 long-only crypto momentum basket that carries the full cycle drawdown. No parameter change fixes a 70% drawdown or turns beta into alpha. Failure pattern: risk_reject/no_edge long-only crypto momentum basket, >50% DD (L19/L45/L49).

Implementation

Long-only cross-sectional momentum on a 15-name Binance USD-M daily universe (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, AVAX, LINK, DOT, ETC, TRX, LTC, UNI, ATOM). Every 30 days on a calendar-anchored grid the universe is ranked by trailing 90-day return on timestamp-aligned closes; the top tercile (5 names) is held long, weighted inverse to each name's 30-day realized volatility so every position contributes equal portfolio variance, at 95% gross exposure with no leverage and no short leg. A -15% per-position stop-loss is checked on every bar independently of the rebalance grid. Pure daily OHLCV — no funding, OI or long/short-ratio data.

Verification Results

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Not the reason for the fail, but for the analyst once the rebalance defect is fixed: the sandbox headline (+935% total return, Sharpe 0.72 CI [0.16, 1.35], PF 1.45) is largely long crypto BETA, not the cross-sectional momentum alpha the hypothesis tests — beta 0.54, benchmark_correlation 0.65 vs the equal-weight basket, information_ratio −0.12, and a 73.7% max drawdown (CI to 89.6%). Long-only top-tercile momentum is highly correlated……Show moreShow less

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Not the reason for the fail, but for the analyst once the rebalance defect is fixed: the sandbox headline (+935% total return, Sharpe 0.72 CI [0.16, 1.35], PF 1.45) is largely long crypto BETA, not the cross-sectional momentum alpha the hypothesis tests — beta 0.54, benchmark_correlation 0.65 vs the equal-weight basket, information_ratio −0.12, and a 73.7% max drawdown (CI to 89.6%). Long-only top-tercile momentum is highly correlated to simply holding the basket; judge alpha vs the equal-weight benchmark and weigh the deep drawdown, not the raw return. A full-backtest question, not a QA block. - The monthly rebalance cadence is driven by a PROCESS-RELATIVE bar counter with the exact -(10**9) sentinel checklist item 18 names as critical. self._last_rebalance = -(10**9) and self._signal_bar_count increments once per calculate_signal; rebalance fires when (self._signal_bar_count - self._last_rebalance) >= self.rebalance_days. Consequences: (a) on the first bar after ANY (re)start, 0 - (-1e9) = 1e9 >= 30, so it immediately does a full 15-name rebalance — and this factory restarts often (Windows DLL degradation, OOM, health-monitor os._exit(99)), so a crash-restart loop churns the whole portfolio and pays full round-trip fees each restart; (b) the monthly grid is anchored to process uptime, not calendar time, so it drifts off any fixed date grid and walk-forward/parity replay can never reproduce the rotation dates. The single-run backtest is deterministic so sandbox numbers aren't corrupted, but this is a real live/paper and replay-parity correctness defect.

Backtest Review

Sharpe: 0.65
Total return: 401.27%
Max drawdown: 70.22%
Trades: 170
Win rate: 37.6%
Profit factor: 1.20

Faithfully implements the post-mortem fixes (long-only, inverse-vol, -15% stop, wider universe); removes the -98% short-leg alt-season blowup

Clean calendar-anchored, restart-safe rebalance; low fee drag (0.86% of gross); adequate 170-trade sample over 6.6 years

max_drawdown 70.2% (CI to 91.4%) — well past the L19 50% hard-abandon line; brutal stress windows (rate-shock -48%, Luna -30%, FTX -33%)

No real edge over the basket: alpha 0.08, beta 0.47, benchmark_correlation 0.68, information_ratio -0.44 — levered alt-cycle beta, not alpha

Sharpe 0.655 with CI [-0.144, 1.412] straddling zero; huge year swings (2022 -65%, 2024 +209%, 2025 -32%)

Headline inflated: end_unrealized_pct 193.9 (~half of 401% is open-position MTM) plus a +45.5% final-day outlier; avg_trade_return_pct 13.2% and exposure 436% are leverage/compounding artifacts (L36)

L45/L49 long-only crypto momentum basket carrying the full cycle drawdown

Outcome Summary

This strategy carefully implemented all four of the analyst's post-mortem recommendations from the abandoned long-short momentum strategy — going long-only to remove the alt-season short blowup, using inverse-vol risk-parity sizing, adding a -15% per-position stop, and widening to a 15-name universe with a 90-day formation window. It reached the backtesting stage, but the engine retrieved 0 daily bars for an ATOMUSDT.BINANCE leg against the required minimum, and after 5 retries the run was abandoned as a data-availability problem rather than a strategy defect. Because the blocker was the data layer, no backtest, optimization, or analyst evaluation ever occurred, and the abandon note flags it for re-evaluation once collection is fixed or the symbol is backfilled. The redesigned momentum thesis therefore remains untested rather than disproven.

A multi-instrument portfolio is only as runnable as its weakest leg's data; confirm every symbol in the basket is collectible/backfilled before committing, since a single missing leg (here an ATOM daily leg) aborts the entire backtest as data_unavailable.

It was abandoned at the data/backtesting stage due to a data-layer blocker, not a strategy fault — backtest data could not be collected after 5 retries — so the backtest-review, optimization, analyst, and risk stages were never reached.

A long-only, OHLCV-only cross-sectional momentum portfolio over a 15-symbol Binance USD-M universe that monthly ranked names by 90-day trailing return, held the top tercile (5 names) with inverse-volatility (risk-parity) sizing and a per-position -15% stop-loss — a faithful re-implementation of the analyst post-mortem fixes for the earlier long-short tercile strategy that blew up -98% in 2021 alt-season.

No performance exists: the backtest never ran because required daily bar data could not be collected (the data layer reported 0 bars for an ATOMUSDT.BINANCE 1-day leg against a needed minimum of 50 / 90+ days span). No metrics, optimization, or analyst review were produced.