Strategies

ETH Single-Asset Time-Series Momentum Long-Short on BINANCE Daily (Second-Asset Validation of BTC TSM, Moskowitz-Ooi-Pedersen 2012, Single Instrument)

Outcome: Abandoned

EthDailyTimeSeriesMomentumVolNormalizedLS

Outcome Summary

This strategy aimed to validate the single-asset time-series-momentum mechanism on a second major, ETH, mirroring the BTC and SOL TSM siblings with a vol-normalized tanh trend signal on safe daily data. Its iteration-2 fix successfully raised trade density to 88 balanced long/short trades, but the base backtest revealed a break-even, outlier-carried edge: a 145% headline return concealed a 0.318 Sharpe with a negative confidence-interval floor, an 80.7% drawdown, kurtosis of 498, and three straight losing years, with 95% of the return sitting in unexited open MTM. The analyst abandoned it at the pre-optimization review, judging that a best-of-225 sweep would merely chase the 2021 outlier and fail the deflated-Sharpe gate, and that the weakness was intrinsic to the mechanism and asset rather than a code defect.

Strategy report

Hypothesis

Second-asset validation of the single-asset time-series momentum mechanism class, deliberately mirroring my in-pipeline BtcSingleAssetTimeSeriesMomentumLongShort hypothesis on ETHUSDT.BINANCE. The analyst's strict refined heuristic explicitly states: 'True validation requires... (b) replicating positive expectancy on a SECOND instrument with the same implementation.' With the BTC single-asset TSM hypothesis just accepted on iter-1 (cleanest verification path in the factory), the natural next step is to test whether the underlying factor (vol-normalized 30-day trend) generalizes to a SECOND major. ETH is the natural choice: similar liquidity tier to BTC, similar academic literature coverage (Hutchinson-O'Brien 2014, Liu-Tsyvinski-Wu 2021 both document TSM on ETH), but DIFFERENT underlying driver (smart-contract ecosystem vs monetary-asset thesis). The single-asset scope means NO multi-asset bar dependency — the 4H/8H translation curse that killed 4 of my prior multi-asset hypotheses can't bite here because there's only ONE instrument. ETHUSDT.BINANCE at daily is on the system's explicit safe-list ('BTCUSDT/ETHUSDT/SOLUSDT.BINANCE on a 1H/4H/1D timeframe is always safe'). Helps direction quota (currently 89.9% long-only — adds long/short/flat positioning). Distinct from BTC TSM because different underlying. CRITICAL DIFFERENTIATORS: (1) ETH has historically had LONGER trend persistence than BTC (ecosystem-narrative cycles last weeks-months), suggesting TSM may show HIGHER Sharpe on ETH than BTC, (2) if BTC TSM validates and ETH TSM ALSO validates, the mechanism class is established as a CLASS — analyst can confidently extend to SOL single-asset TSM next, (3) if BTC TSM works but ETH TSM doesn't, mechanism is BTC-specific (similar to how Donchian transferred only partially across assets in prior analyst findings).

Iteration 2 addresses the analyst's trade-density diagnosis: the prior t-stat>=1.5 hard gate fired only 34 times in ~6 years (~6/yr, below the ~62 walk-forward floor) and produced an outlier-dominated sample (kurtosis 1398) that would fail the deflated-Sharpe gate. Per the explicit feedback, I aligned the signal construction with the SOL single-asset TSM sibling that reached 261 trades / Sharpe 1.01: the raw OLS trend t-stat is now squashed through tanh(t/tanh_scale) into a continuous bounded [-1,1] score, and entry uses a low tanh threshold (0.25) with a 0.10 exit band instead of the sparse |t|>=1.5 gate. I also shortened the trend lookback 60->25 to make the trend estimate more responsive (faster exits, shorter holds), further lifting trade density. The vol-target sizing, wide safety stop, and OLS trend core (all of which passed Layers 1-3) are unchanged, and the tanh scale keeps the signal continuous and finite so the synthetic-scenario and static layers remain green. This keeps the same long/short TSM mechanism as the BTC/SOL siblings for a genuine second-asset validation while targeting trade_count > 60 with a de-lumped return distribution.

Not worth optimizing: sub-significant, outlier-carried base. profit_factor 1.0445 is break-even after fees; return_kurtosis 498 / tail_ratio 0.165 means the +145% return hinges on ~4 single days (2021-02-26 +142.8% chief among them); sharpe_ci_low -0.474 and PSR 0.875 fail significance BEFORE any multiple-testing deflation; max_drawdown 80.7% with three consecutive losing years (2022/2023/2024). This is the same outlier-luck signature repeatedly abandoned at the deflated-Sharpe gate this session for single-asset daily trend/momentum on majors (the ETH ATR-trail trend-follower — a stronger base than this — still failed at DSR 0.78). Best-of-225 optimization would chase the 2021 outlier and predictably fail DSR; the weakness is intrinsic to the mechanism/asset, not a code bug (the developer already iterated once to fix trade density — count is now fine, quality is not). Saving the 2-hour optimization cycle.

Implementation

Single-asset time-series momentum (long/short/flat) on ETHUSDT.BINANCE daily bars. Fits an OLS log-price trend over a 25-day lookback, converts the slope t-stat into a bounded [-1,1] trend score via tanh(t/scale), and enters long/short when |score| >= 0.25, exiting when the score decays into a 0.10 band, flips, or a wide 20% safety stop triggers. Positions are vol-targeted (target_vol/realized_vol, capped at 1x equity). Pure OHLCV, single instrument, leverage 1.0.

Backtest Review

Sharpe: 0.32
Total return: 145.44%
Max drawdown: 80.70%
Trades: 88
Win rate: 35.2%
Profit factor: 1.04

Signal functions correctly: 88 trades (42 long / 46 short) over 6yr, balanced long/short as the TSM hypothesis intends, 0 entries dropped — the iteration-2 trade-density fix worked

Single-instrument daily ETH is on the safe data path; no multi-asset translation risk

profit_factor 1.0445 — edge is essentially break-even after fees (~1.23% of gross), the decisive red flag

return_kurtosis 498 / skew 18.4 / tail_ratio 0.165 — outlier-carried; the +145% return hinges on ~4 single days led by 2021-02-26 (+142.8%)

max_drawdown 80.7% (CI to 99.4%), 8 consecutive losses — catastrophic risk

sharpe_ci_low -0.474 and PSR 0.875 — base Sharpe not significant even BEFORE the 225-trial deflation penalty; a deflated-Sharpe gate would reject it

Three consecutive losing years (2022 -26%, 2023 -30%, 2024 -77%); recent gains are a few outlier trend days

end_unrealized_pct 95.3 — the headline is almost all un-exited open MTM, not realized edge

Analysis

Do not optimize — the sample is far too thin and outlier-dominated (identical issue to the BTC TSM sibling, slightly worse here). Only 34 trades over ~6 years (~6/yr), below the ~62 walk-forward floor: the t-stat entry (entry_t 1.5, lookback 60) plus 62-day average holds fire too rarely, and the +1148% headline rests on ~2 trade-close lumps (2021-06-11 +255%) plus a 359%-of-headline unrealized open position (kurtosis 1398, skew 35.5). A 225-trial optimization on 34 lumpy trades will fail the deflated-Sharpe gate (the 37-trade DOGE breakout came back DSR 0.0). FIX — increase trade density to >60-100 while keeping the long/short TSM mechanism: lower entry_t (test ~0.75-1.0), shorten the trend lookback (test 20-30 instead of 60), and/or shorten holds. CRITICAL: align the signal construction with the SOL single-asset TSM sibling, which achieved 261 trades and optimized to Sharpe 1.01 using a tanh-threshold signal (entry_threshold 0.25) rather than the t-stat>=1.5 gate — the t-stat formulation is what starves both the BTC and ETH versions. Apply the SAME fix to the BTC sibling for consistency. Re-backtest and confirm trade_count > 60 AND return_kurtosis is reasonable (no single trade dominating) AND a positive sharpe_ci_low before resubmitting. If a denser config still cannot exceed ~60 trades or remains outlier-dominated, abandon — single-asset t-stat TSM on ETH daily is then too sparse/lumpy to validate, and the mechanism is carried by the SOL sibling.

Outcome Summary

Fixing trade density (the iteration-2 change that lifted the count to 88) does not fix trade quality: single-asset daily trend/momentum on majors keeps producing the same outlier-luck signature that break-even profit factor and extreme kurtosis expose before optimization can help.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate, so the strategy never reached optimization; the decisive issues were a break-even edge (profit factor 1.045 after ~1.23% gross fees), a catastrophic 80.7% drawdown (CI to 99.4%), and a sub-significant, outlier-carried Sharpe whose +145% return hinged on roughly four single days led by 2021-02-26 (+142.8%).

A second-asset validation of single-asset time-series momentum (Moskowitz-Ooi-Pedersen 2012) on ETHUSDT.BINANCE daily bars, using a vol-normalized log-price trend t-stat squashed through tanh into a bounded long/short/flat score with inverse-volatility (vol-target) position sizing.

The base backtest produced 88 balanced trades (42 long / 46 short) over ~6 years with a headline total return of 145.4% but a Sharpe of only 0.318 (sharpe_ci_low -0.474), profit factor 1.045, and an 80.7% max drawdown; return kurtosis was 498 with skew 18.4 and end_unrealized_pct 95.3%, indicating the gains were outlier-driven and mostly unrealized open MTM. It had three consecutive losing years (2022 -26%, 2023 -30%, 2024 -77%).

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