Time-series momentum is the most replicated anomaly in the academic literature, and after two years of testing everything else, it is also the only signal my own research pipeline kept validating. This post documents the broadest single test: one fixed rule, applied unchanged, across three asset classes.
The rule
Long-only trend state, decided at each daily close:
- Price filter: close > 200-day simple moving average
- Momentum filter: trailing 60-day return > 0
- Position taken at the next open, exited the same way; flat otherwise
No fitted model, no parameters selected on data — the windows come from the original crypto study that produced this rule family and were frozen before this test. Fees: 5bps/side (liquid-ETF spread + impact), stress-tested at 15bps + half-spread.
The basket
28 liquid ETFs spanning broad equities (SPY, QQQ, IWM), international (EFA, EEM), nine sector funds, homebuilders/biotech/semis, treasuries and credit (TLT, IEF, LQD, HYG, TIP, VNQ), metals (GLD, SLV, GDX), energy (USO, DBC) and the dollar (UUP). Twenty years of daily data.
Results
The headline cell — net-positive on 28 out of 28 assets:
| rule | median net | mean net | positive / total | both halves |
|---|---|---|---|---|
| dual 200/60 | +7,189 bps | +8,242 bps | 28 / 28 | 20 / 28 |
| dual 200/120 | +7,168 bps | +10,065 bps | 28 / 28 | 24 / 28 |
340 of 448 grid cells were net-positive on both independent halves of the sample. The champion cells survived a taker-fee stress test with thousands of basis points of margin — holds last weeks, so fees amortize over whole trends.
The honest part: what the null says
A portfolio combining all 28 sleeves (equal-weight, then inverse-volatility weighted) delivered Sharpe 0.63–0.78 across neighboring rules, with max drawdown around −8 to −12% versus −54% for buy-and-hold over the same period. Robustness checks passed: leave-one-out Sharpe range [0.63, 0.68] across all 28 drops, no single-asset dependence.
But the permutation null matters. When asset returns are circularly shifted under the fixed position streams — destroying timing while preserving exposure — random books earn Sharpe ≈ 0.57 against our realized 0.66 (p = 0.21). Translation: most of the portfolio's return is secular bull-market drift captured by long-only trend exposure, not timing skill. This matches what the Monte-Carlo analysis of the original single-asset version showed: it is beta capture with a filter.
That is not a knock-out. The product here is drawdown transformation — equity- class participation at a quarter of buy-and-hold's pain, plus genuine crash convexity (the book sits flat or short nothing through bears by construction). But it should be sized and marketed as what it is.
Portfolio construction notes
Inverse-volatility weighting across active sleeves improved risk-adjusted returns (Sharpe 0.75, MaxDD −8.4%). Volatility targeting to 10% annualized with a 2× cap added CAGR without proportional drawdown in backtests. Concentration — keeping only the top-K strongest trends — consistently hurt: breadth is where the stability comes from.
Deployment
This system now runs as a live paper sleeve: daily decisions on completed bars, 16:20 ET, orders filling next open, idempotent per-date state, double-gated order submission. Its first weeks are tracked publicly on this site's dashboard.
The full grid CSV, per-cell ledgers, and the permutation harness are available in the research repository.