XSMOM — Cross-Sectional Momentum

Core idea

In crypto, relative momentum persists over multi-week horizons: coins that have outperformed on a risk-adjusted basis tend to keep outperforming, and the weakest tend to keep underperforming. Ranking by vol-normalised return (return ÷ volatility) rather than raw return avoids buying high-volatility noise — a coin that rose 30% on 60% daily vol is a weaker signal than one that rose 20% on 15% vol.

Why the long filter exists

The vol-normalised metric can rank a coin highly because its volatility collapsed, not because it's trending up. The long side therefore requires the coin to be above its moving average (MA-100 after optimisation) and have a positive raw 30-day return. The short side needs no filter: in a structurally risk-off altcoin regime, the weakest losers keep losing.

Signal — vol-normalised return (VNR)

For each symbol, over the trailing 30 days ending at the rebalance date:

raw_return   = (close[t] - close[t-30]) / close[t-30]
atr_pct      = ATR(14) / close[t]            # daily vol, as % of price
VNR          = raw_return / (atr_pct × √30)  # risk-adjusted momentum

VNR measures how far price moved per unit of volatility over the window — the standard "vol-normalised" momentum used in academic cross-sectional momentum studies.

Selection rules

• Longs (top 5 by VNR) — must pass trend filter: close > MA(100) and raw_return > 0. Only coins with genuine upward momentum qualify.
• Shorts (bottom 5 by VNR) — require VNR ≤ -0.5 (optimised floor). Important finding: with this floor, shorts rarely trigger in this sample — the optimised book is effectively long-only. When shorts are forced (no floor), they lose money over the full history (see Monte Carlo section).
• Sizing — long weights are linear in VNR (strongest signal → biggest position); shorts are equally weighted.
• Exit — positions close at the next rebalance, or earlier at a 15% stop-loss.

Headline metrics (104 periods)

Per-year P&L

The strategy is a momentum book that needs a bid: it loses in the 2022 bear market and the 2021 sample, and compounds in 2023–2026. Expect underperformance in sustained downtrends — this is regime-dependent, not regime-neutral.

Jackknife — single-period dependence

We remove each period in turn and recompute total P&L. The most influential periods:

The original 21-period backtest had a single week worth 91% of P&L. After extending to 104 periods, the best week is 22.7% — removing it leaves +$4,926, still strongly positive. No single period breaks the result.

Block bootstrap — outcome distribution

20,000 simulated paths, each a resample of the period-return series in blocks of 3 (preserves short-run autocorrelation / regime clustering):

Median path +$5,381 (close to the observed +$6,371); the entire P5→P95 range is positive. P(total > 0) = 99.8% across all 20,000 paths.

Robustness matrix — does the config matter?

Four very different parameterisations all end positive over 4.75 years — the edge is not an artifact of one lucky config.

Walk-forward split-half (OPT config)

The edge strengthened sharply in the second half (Sharpe 0.13 → 0.34). The first half includes the 2022 bear (where the book loses) and the early deep-history sample. This concentration is the honest caveat: the strategy's best years are the most recent ones, so forward-looking expectations should be anchored to the second-half economics, not the full-sample average.

Rebalance frequency biggest lever

Extending the hold from 7→14 days lets winners compound and cuts rebalance whipsaw. This leg ran without the short floor, so shorts were still active here.

Stop-loss best Sharpe

15% is the single best Sharpe improvement: it caps catastrophic single-coin blow-ups while letting winners run.

Trend-filter MA period

MA=100 is the sweet spot — a slower regime filter that rejects more false-positive longs in chop.

Short VNR floor

The floor improves PF by culling weak/mean-reverting shorts — but note the robustness finding: over the full 4.75y sample, forcing shorts on (no floor) is a net drag (-$205 over 750 trades). The floor's value is largely "don't short at all."

Signal window (VNR days)

30 days has the best Sharpe / drawdown balance. 14d has higher raw P&L but 2.9× the drawdown.

Portfolio size & concentration

Concentration caps hurt: signal-proportional sizing is an alpha source, not a risk bug.

Repository layout

MOMSXperp/
├── config.py         # strategy parameters (optimised)
├── run.py            # backtest entry point
├── analysis.py       # ranking, selection, sizing, P&L engine
├── fetch.py          # Bybit V5 OHLCV fetch + paginated deep history
├── refetch_deep.py   # re-fetch deep history for all symbols
├── report.py         # extended diagnostics
├── charts.py         # matplotlib equity/side/drawdown charts
├── sweep.py          # 8-dimension parameter sweep
├── montecarlo.py     # jackknife + block-bootstrap robustness
├── robustness.py     # multi-config matrix + walk-forward split
├── xsmom_results.json       # optimised backtest (104 periods)
├── xsmom_mc_results.json    # Monte Carlo output
├── xsmom_robustness.json    # robustness matrix
├── xsmom_sweep_results.json # full sweep output
└── site/             # this Cloudflare Pages site
    └── assets/       # charts

Run it yourself

# backtest (cached data; --force to re-fetch deep history)
python3 -m xsmom.run

# deep re-fetch of all symbols (one-time)
python3 -m xsmom.refetch_deep

# Monte Carlo robustness (20k paths)
python3 -m xsmom.montecarlo

# multi-config robustness matrix + walk-forward
python3 -m xsmom.robustness

# extended diagnostics / charts
python3 -m xsmom.report
python3 -m xsmom.charts

# full parameter sweep (2-3 min)
python3 -m xsmom.sweep

Data is cached under xsmom/data_cache/ so repeat runs are instant. No API key required for public market data.