The Momentum Mirage: +106% One Year, −43% the Next

Verdict: overfit / regime-specific · Asset/TF: 18 cryptocurrency futures, weekly · Sample: 2021-01 to 2026-08

This experiment finally escaped the transaction-cost trap.

The strategy traded once a week. Fees averaged only 0.063% per week. Funding cost another 0.040%. Together they were small relative to the underlying price moves.

And the strategy still failed spectacularly.

Weekly cryptocurrency momentum returned +106% in 2025 and then lost 43% in the frozen 2026 test. The signal itself reversed. No fee assumption could explain it away.

The paper behind the test

Ao Yang (2025) studies risk-managed momentum in cryptocurrency markets. The paper forms a conventional Winner-minus-Loser portfolio using a two-week formation period and one-week holding period, then scales momentum exposure according to forecast risk.

The reported result is attractive: average weekly return rises from 3.18% to 3.47%, while annualized Sharpe improves from 1.12 to 1.42.

We implemented a tradable futures adaptation with two questions:

1. Does conventional weekly crypto momentum survive fees and funding?
2. Does inverse-volatility scaling improve it out of sample?

Hypothesis

Every Sunday UTC, rank cryptocurrencies by their cumulative return over the previous 14 days. At Monday open:

• long the highest-return quintile;
• short the lowest-return quintile;
• allocate 50% gross exposure to each side;
• hold until the next Monday open.

The fixed universe contained 18 futures: BTC, ETH, BNB, XRP, ADA, DOGE, SOL, DOT, LTC, LINK, BCH, ETC, TRX, XLM, EOS, AVAX, DASH and ZEC.

Risk-managed variants scaled the entire portfolio using only previously completed weekly WML returns. Volatility lookbacks were frozen at 4, 8 and 12 weeks; exposure was clipped between 0.25× and 2.0×.

Strategy

{
  "formation_days": 14,
  "holding_days": 7,
  "signal_time": "Sunday UTC close",
  "execution": "next Monday open",
  "entry_rules": [
    "long equal-weight top return quintile",
    "short equal-weight bottom return quintile"
  ],
  "risk_lookbacks_weeks": [4, 8, 12],
  "scale_bounds": [0.25, 2.0],
  "costs": "5.4044 bps one-way × actual L1 turnover",
  "funding": "historical Binance 8-hour funding with long/short sign"
}

First result: risk management made momentum worse

Train covered 2021–2023. Validation covered 2024–2025. Target volatility was fixed from Train, and Validation could choose exactly one lookback only if its net Sharpe exceeded plain momentum.

The 12-week scaler modestly improved total return and drawdown, but it lowered Sharpe. The shorter lookbacks increased turnover and performed worse on every primary risk-adjusted comparison.

The paper's central claim therefore failed our Validation gate. The risk-managed parent hypothesis was rejected without opening 2026.

The baseline earned a separate Final test

Plain momentum was not the original winner we wanted, but it had positive net expectancy and Sharpe after fees and funding. Because 2026 had not yet been calculated, we registered a child hypothesis before opening it:

• preserve the exact universe and weekly timing;
• remove volatility scaling;
• prohibit leverage, stops, regime filters and retuning;
• evaluate 2026 once.

That frozen test failed every economic gate except sample size.

Both temporal halves were negative: −2.52% per week in the first half and −0.87% in the second.

This was not a fee problem

Average weekly Final attribution was:

More than 93% of the mean loss came from the price signal. Reducing commissions to zero would not rescue the strategy. Nor would ignoring funding.

This is fundamentally different from our minute-level lead–lag experiments, where a positive statistical edge was buried under turnover. Here turnover was manageable; the forecast direction stopped working.

The aggregate Validation hid a regime bet

The yearly path explains the failure:

The apparently healthy 2024–2025 Validation result was almost entirely a 2025 phenomenon. Aggregating two years produced +85.5% and Sharpe 0.895, while one constituent year was already negative.

Our preregistered gate required positive aggregate Validation performance, but not positive performance in each calendar subwindow. That omission did not create look-ahead in the Final result, but it allowed a regime-specific strategy to advance.

What risk management can and cannot do

Inverse-volatility scaling changes how much capital is allocated. It does not determine whether past winners will continue outperforming past losers.

If high volatility predicts momentum crashes, scaling down can help. In our sample, trailing WML volatility did not reliably identify the periods when cross-sectional momentum would reverse. The scaler slightly softened one drawdown but did not stabilize performance across regimes.

The lesson is broader than momentum:

Risk management can reshape a valid edge. It cannot manufacture persistence in an unstable signal.

Limitations

• The universe contains currently surviving futures and therefore has survivorship bias.
• Portfolios are equal-weighted rather than market-cap-weighted as in the source study.
• Futures prices and realized funding make the test executable, but differ from the paper's CoinMarketCap spot universe.
• The 2026 corpus has been used by unrelated research families, limiting the maximum possible verdict.
• Thirty-one Final weeks are useful but not enough for institutional promotion.

These limitations weaken any positive claim. They do not explain a −43% frozen result with both halves negative.

Research trail

Tools called: get_ai_instructions → create_hypothesis → Binance daily futures and funding download → weekly WML construction → volatility scaling → chronological Validation → child hypothesis preregistration → frozen Final → ai_run_backtest → update_hypothesis → create_canvas_note.

Risk-managed parent: hyp_1786510275644_23 · Plain child: hyp_1786510531371_25 · Both verdicts: REJECTED.

Reproduce

Primary pipeline: research/paper_weekly_risk_managed_momentum.py. Frozen child test: research/paper_plain_weekly_momentum_followup.py. Results live under research/results/paper_weekly_risk_managed_momentum/ and research/results/paper_plain_weekly_momentum_followup/.

This experiment finally found an edge large enough to survive fees. It revealed the harder problem: surviving the next regime.