Skip to content

cross_sectional_momentum.md

Latest commit

 

History

History
94 lines (76 loc) · 4.3 KB

cross_sectional_momentum.md

File metadata and controls

94 lines (76 loc) · 4.3 KB
Raw

cross_sectional_momentum — Top-N rotation across a universe

Cross-sectional momentum (CSMOM) is a portfolio-level strategy: each week, rank the universe by trailing N-day return and hold the top-K. The Research-Claude survey lists CSMOM at priority 5/5 citing:

  • Liu, Tsyvinski, Wu (2022). Common Risk Factors in Cryptocurrency. Journal of Finance 77(2):1133-1177. Identifies market / size / momentum as the three factors that explain the cross-section of cryptocurrency returns and reports ten significant long-short strategies.
  • Tzouvanas, Kizys, Tsend-Ayush (2019). Momentum trading in cryptocurrencies: short-term returns and diversification benefits.
  • Starkiller Capital — practical reproduction with a 15-35 day lookback and a 7-day rebalance.

Because the strategy makes a portfolio-level decision over multiple assets simultaneously, it does not fit the single-asset Strategy interface. It lives in trade_lab/backtest/cross_sectional.py and is called directly via run_cross_sectional_momentum(...).

Rules

  1. Universe. The caller supplies a dict mapping symbol → OHLCV DataFrame. All assets are outer-joined onto a common date index; missed candles inside each asset's listed history are forward-filled, but we never invent pre-listing prices.

  2. Trailing-return ranking. Every rebalance_days (default 7), compute close.pct_change(lookback_days) per asset (default 30-day lookback). The decision uses only closes through the rebalance date.

  3. Selection. Keep only assets with positive trailing return; from those, take the top top_k (default 3).

  4. Optional BTC regime gate. Pass btc_candles= to add a BTC > SMA(btc_gate_sma_period) filter (default 200). When the gate is closed, the portfolio sits in cash that week regardless of the cross-section.

  5. Weighting.

    • equal (default): 1/len(basket) per selected asset.
    • inverse_vol: weight by 1 / realized_vol(vol_lookback), normalized to sum to 1. Empty / zero vol assets fall back to equal weighting within the basket.

  6. Execution. Target weights from rebalance date N apply to bar N+1 onward, exactly like the single-asset engine. Costs are (fee_rate + slippage_rate) * turnover summed across assets.

Why long-only

The Liu-Tsyvinski-Wu (2022) original is long-short. We strip the short leg because spot Binance does not provide it; the long-only retail analogue is what the Research-Claude survey calls out as priority 5/5. Expect lower Sharpe than the academic figure (which is a long-short factor return).

Output

CrossSectionalResult carries:

Field Description
equity Portfolio equity curve.
returns Net per-bar returns (after fees + slippage).
weights DataFrame: rows = dates, cols = assets, values = held weights ∈ [0, 1].
rebalance_dates Dates where the weight vector changed.
total_return Final / initial – 1.
max_drawdown Peak-to-trough on equity.
sharpe Annualized Sharpe (using annualization_factor).
num_rebalances Count of weight-vector changes.
average_basket_size Mean number of held assets per bar.
average_cash_fraction Mean 1 - sum(weights) per bar.
total_fees, total_slippage Dollar costs.

Important caveats

  • Survivorship bias. A real-world CSMOM in 2020 would have included LUNA, FTT, AAVE/early DeFi, etc. — several went to zero or near-zero. Running on a hand-picked surviving universe (BTC/ETH/BNB/SOL) gives a flattering result.
  • Small universe. The literature suggests 10-30 coins. With four, top-2 selection is highly concentrated; the result is more like "rotation between two of four" than a cross-section.
  • Fees. Weekly rebalance × small basket can compound costs quickly on a small account; see total_fees in the result.
  • No vol scaling at portfolio level. Each asset's vol affects the inverse-vol weights, but the total exposure is whatever weights sum to. Pair this with weighting='inverse_vol' to dampen the portfolio vol implicitly.

See docs/results/strategy_comparison.md for the side-by-side results.