MOMSXperp — XSMOM Backtest Review

02What was wrong with the earlier analysis

Each item below changed the numbers; together they explain the gap between the old and new results.

Look-ahead and accounting

Stop-loss used the future: the old engine capped each trade's end-of-period loss at −15%, a free put that never stopped out trades that dipped and recovered. On the old data it added $1,377 of $6,371.
Published files were stale: the engine was rewritten but README, site and JSON were never regenerated.
ATR overstated 30/14×: 30 true ranges summed then divided by 14; the short floor −0.5 really meant −1.07.
Sizing: min–max scaling gave the weakest long zero weight (a phantom "losing" trade) and old gross exposure swung from $100 to $1,614 per period.
Sharpe annualised with √52 on biweekly periods (should be √26).

Universe and testing design

Survivorship / look-ahead universe: today's top-50 by 24h volume, re-downloaded every run (non-reproducible, and it included gold, oil and a 3× equity ETF).
Parameters chosen on 24 periods (~5 months) with the look-ahead stop, weekly vs biweekly compared over different spans.
Robustness matrix leaked state (BOTTOM_N=0 carried into the next config) and compared configs over different date ranges.
"Walk-forward" was a split-half of in-sample-optimised parameters; Monte Carlo resampled an already-biased P&L; no fees or funding.

03Results on the corrected engine

Cumulative return of the headline config, the original design, the survivorship-biased replica and BTC buy-and-hold

Cumulative return as % of a $100 capital base, non-compounded, after fees, slippage and funding.

Configuration	Periods	Sharpe	Ann. return	Max DD	Long P&L	Short P&L	Trades

By year (% of capital)

Original design P&L decomposed into long price, short price, funding and costs

The original design's P&L split into price moves, funding and trading costs. Funding received on longs in squeezed small caps is the largest single source.

Drawdown in % of capital (gross exposure up to 200%).

03bComposition and rotation (headline config)

Is the long book rotating, or a proxy for holding a few coins? Computed from the per-trade records by composition.py.

Rotation

Top contributors (cumulative long P&L, $)

Symbol	Periods	Wins	Cum P&L

Long book by rebalance

Rebalance	Long book

04Validation tests

Every test uses the same universe and cost model. "Headline" is the sweep-optimised config; "Original" is the pre-sweep design (weekly, MA50, no stop, bottom-5 shorts), which the full-history sweep ranks much higher.

Test	Headline	Original	What it tells you

Execution and accounting sensitivity (annualised Sharpe)

Variant	Headline	Original

06Monte Carlo (headline config)

Moving-block bootstrap of the realised per-period returns. This measures sampling noise only; it cannot correct selection bias, which is what the deflated Sharpe and PBO above address.

Block length	P(total > 0)	Sharpe P5	Sharpe P50	Sharpe P95	Max DD P95

07Methodology

Data

Crypto only: stock, ETF, commodity and forex perps, stablecoins and tokenised gold are excluded.
Daily candles with turnover; the in-progress candle is dropped. Funding from Bybit's settlement history for every symbol that ever entered the universe.

Engine

Universe at each rebalance = top 50 by trailing 30-day turnover among coins trading that day.
Signal = 30-day return / (mean daily true range % × √30). Entry at the signal candle's close; exit at the next rebalance close; intraday stops fill at the stop or the gapped open.
$100 long budget (weights ∝ signal) and $100 short budget (equal weight) on $100 capital. Taker fee 0.055% + 0.05% slippage per side, charged only on traded notional. Delisted coins exit at their last price.

08Caveats

Funding is the fragile part. Much of the better configs' P&L is negative funding paid to longs in squeezed small caps. That is capacity-limited, crowded and depends on getting filled in illiquid books; 0.05% slippage is optimistic there.
Returns are non-compounded on a fixed $100 base with up to 200% gross exposure; drawdowns above 100% of capital mean the book would have been liquidated without de-leveraging.
Daily candles cannot see intraday path order; stops use daily highs and lows.
Every configuration here was still chosen by looking at this history. Treat the best sweep result as a hypothesis for paper trading, not an expected return.

09Reproduce

cd ..                                          # parent of the MOMSXperp package
python -m MOMSXperp.refetch_deep --funding     # snapshot instruments, klines, funding (~10 min)
python -m MOMSXperp.run                        # headline backtest
python -m MOMSXperp.robustness                 # config matrix
python -m MOMSXperp.montecarlo                 # bootstrap + jackknife
python -m MOMSXperp.sweep                      # 360-config grid, deflated Sharpe, PBO
python -m MOMSXperp.validation                 # benchmarks, null, IC, phase, walk-forward
python -m MOMSXperp.validation --config original
python -m MOMSXperp.composition                # rotation / contribution
python -m MOMSXperp.charts
python -m MOMSXperp.site_build                 # refresh this page's data
python MOMSXperp/tests/test_core.py            # unit tests

XSMOM cross-sectional momentum, re-tested

01Verdict