Sharpe 10.73, Still Untradeable: The Cross-Crypto Experiment Gets 16 More Chances

Verdict: no economic edge · Asset/TF: BTC-led 16-altcoin futures, 1m–10m · Sample: 180 days, 17,278 five-minute Validation decisions

Our previous experiment found a real-looking BTC lead–lag signal worth only 0.033 basis points in an ETH–SOL pair. Execution costs were 327 times larger.

There was an obvious objection: two altcoins are not a cross-section.

The academic study behind the idea used 30 cryptocurrencies, forecast sorts, and diversified long–short portfolios. Perhaps ETH and SOL were simply too similar. A wider universe might turn tiny differences in response speed into a meaningful spread.

So we gave the hypothesis 16 altcoins and a second chance.

It worked—statistically. The best gross Sharpe rose to 10.73 and all 16 assets had positive BTC lead coefficients.

It still lost after costs.

The published idea

Guo, Sang, Tu and Wang (2024) report cross-cryptocurrency return predictability using minute-level Binance data. Bitcoin often reacts first to common shocks; returns of other cryptocurrencies respond with a lag. The authors use adaptive LASSO, PCA and BTC-based forecasts to sort coins into portfolios, reporting positive long–short spreads out of sample.

Our first simplified replication ranked only ETH against SOL. This follow-up moved closer to the paper's portfolio construction while keeping the rules simple enough to audit.

Hypothesis

BTC's trailing return should forecast heterogeneous next-period returns across altcoins. If information diffuses at different speeds, the highest forecast quintile should outperform the lowest forecast quintile.

The fixed universe contained BTC as predictor and 16 targets:

ADA, APT, ARB, AVAX, DOGE, ETH, INJ, LINK, LTC, NEAR, OP, SEI, SOL, SUI, TIA, WIF.

For every target, we fit a Train-only regression:

next altcoin return = alpha + beta × trailing BTC return

At each decision, the portfolio went long the three highest forecasts and short the three lowest forecasts, with 50% gross exposure on each side.

Strategy

{
  "predictor": "trailing BTCUSDT return",
  "targets": 16,
  "candidate_horizons_minutes": [1, 5, 10],
  "entry_rules": [
    "long the three highest Train-fitted forecasts",
    "short the three lowest Train-fitted forecasts"
  ],
  "position_weights": "equal within each side; 50% long + 50% short",
  "rebalance": "non-overlapping, once per holding horizon",
  "cost_model": "5.4044 bps one-way × L1 portfolio turnover",
  "selection": "at least 10/16 positive betas and positive Validation net EV and Sharpe"
}

Transaction costs followed actual weight changes. If a coin remained in the same portfolio at the same weight, the model did not pretend to close and reopen it. Initial entry and terminal liquidation were charged.

The signal became stronger

The broad portfolio materially improved the gross result:

That is approximately five times more gross expectancy and an order-of-magnitude improvement in the execution gap.

The economic intuition also became cleaner. At five minutes, all 16 Train coefficients on lagged BTC returns were positive. This is the cross-sectional signature one would expect if BTC processes common information before the altcoin universe.

Validation results

The five-minute implementation generated 17,278 independent decisions across 60 Validation days. Average portfolio turnover was 0.829 per rebalance.

Its break-even execution requirement was severe:

• break-even one-way cost: 0.203 bps;
• break-even full round trip: 0.406 bps;
• modeled full round trip: 10.809 bps.

The strategy would need institutional maker rebates, unusually low market impact and near-perfect infrastructure merely to approach zero. Even a four-basis-point full-roundtrip assumption would remain roughly ten times too expensive for the observed gross edge.

Why a double-digit Sharpe is not enough

The gross portfolio is consistent because it trades constantly. That regularity compresses daily variance and produces a striking annualized Sharpe.

But high-frequency Sharpe can hide bad unit economics. Before costs, thousands of tiny positive observations add up smoothly. After costs, every rebalance introduces a deterministic negative term much larger than the forecast spread.

This is why a research platform should store at least three separate quantities:

1. predictability: is there a stable conditional relationship?
2. tradability: does the spread exceed turnover-adjusted costs?
3. capacity: would the edge survive latency and market impact at useful size?

This experiment produced evidence for the first and failed the second. Capacity therefore never became relevant.

Robustness and the test we did not open

• The 180-day corpus was split chronologically into 60 days Train, 60 days Validation and approximately 60 days Final.
• Horizons, universe, quintile width, cost model and gates were registered before returns were calculated.
• Holding periods did not overlap.
• Only a completed BTC return entered each forecast.
• No horizon achieved positive net Validation expectancy.
• Therefore the pipeline did not read the Final rows.

The result has two important limitations. The 17-asset universe was fixed from currently available files, creating survivorship bias, and the corpus had been used by unrelated strategy families. Even a successful Final would therefore have been capped at WATCH, not promoted.

What changed from the first experiment?

The broader cross-section did exactly what we hoped: it created more forecast dispersion and increased gross alpha. That is useful evidence about the paper's mechanism.

But it also revealed the likely scale required for implementation. Going from two to 16 targets improved the execution gap from 327× to 26.6×. Closing the remainder would require a much richer universe, better forecasts, dramatically lower turnover, or all three.

The next scientifically valid version needs:

1. a point-in-time universe to remove survivorship bias;
2. the paper's broader adaptive-LASSO or PCA forecasts;
3. venue-specific maker/taker and rebate schedules;
4. queue-position and market-impact estimates;
5. new untouched data.

Until those inputs exist, optimizing the current 180 days would turn a clean negative result into a backtest artifact.

Research trail

Tools called: get_ai_instructions → create_hypothesis → synchronized futures panel → Train-only clustered regressions → forecast quintiles → turnover cost audit → report_research_event → update_hypothesis → create_canvas_note.

IDEA MAP: hyp_1786509851298_19 · Verdict: REJECTED · Final rows read: false.

Reproduce

Pipeline: research/paper_broad_cross_crypto_quintiles.py. Frozen metrics: research/results/paper_broad_cross_crypto_quintiles/summary.json. Data: synchronized Binance-style one-minute futures OHLCV files.

The broader experiment did not rescue the strategy. It gave us something better: a quantitative measurement of how far the idea remains from the market.