The 0.033-Basis-Point Edge: When Prediction Is Real but Trading Is Impossible

Verdict: overfit / no economic edge · Asset/TF: BTC leader, ETH–SOL pair, 1m–10m · Sample: 2021-01 to 2026-01, 264,958 one-minute Validation decisions

Suppose a signal has a gross Sharpe ratio of 5.75.

Would you trade it?

Now suppose its average profit is 0.033 basis points, while one realistic round trip costs 10.81 basis points.

That is not a typo. The cost is 327 times larger than the edge.

This experiment produced one of the most useful distinctions in quantitative research: a market can be statistically predictable and still be economically untradeable.

The paper behind the experiment

Guo, Sang, Tu and Wang (2024) study minute-level returns for 30 cryptocurrencies on Binance. They find that lagged returns of other coins predict a focal cryptocurrency, with Bitcoin frequently acting as a leader. Their interpretation is gradual information diffusion: a common shock reaches the most closely watched assets first and slower coins react later.

The paper goes beyond regressions. It forms out-of-sample long–short portfolios from adaptive-LASSO, PCA and univariate BTC forecasts. In the published results, the univariate BTC sort produces a 1.54 bps minute-level spread before the paper's trading implementation, while broader models remain profitable under lower transaction-cost assumptions.

We asked a narrower question:

Can BTC's recent return rank ETH versus SOL strongly enough to trade a market-neutral pair under the costs actually used in our RLXBT research stack?

This is a simplified three-asset translation, not an exact reproduction of the paper's 30-coin portfolios.

Hypothesis

At each decision timestamp:

1. observe BTC's trailing return over h minutes;
2. use Train-only regressions to forecast the next h-minute ETH and SOL returns;
3. go long the higher forecast and short the lower forecast;
4. exit after h minutes;
5. deduct 10.8088 bps per equal-gross pair round trip.

We preregistered h ∈ {1, 5, 10} minutes. All trades were non-overlapping, missing timestamps were rejected, and only Validation could choose a horizon. Final data would be loaded only if the selected horizon had positive net expectancy and both BTC spillover coefficients were positive.

Strategy

{
  "predictor": "trailing BTCUSDT return",
  "targets": ["ETHUSDT", "SOLUSDT"],
  "horizons_minutes": [1, 5, 10],
  "entry_rules": [
    "long the target with the higher Train-fitted forecast",
    "short the target with the lower Train-fitted forecast"
  ],
  "exit_rules": ["exit after the selected non-overlapping horizon"],
  "gross_exposure": "50% long + 50% short",
  "round_trip_cost_bps": 10.8088,
  "selection": "positive Train betas for ETH and SOL, then highest positive Validation net EV"
}

First surprise: the one-minute mechanism partly exists

On Train, the one-minute BTC coefficients were positive for both targets:

SOL reacted more strongly than ETH to the preceding BTC move. That difference gave the pair model something to rank. The result is consistent with a small, fast information-diffusion effect—especially for SOL.

But it did not persist across horizons:

At five and ten minutes, both signs reversed. Whatever is transmitted appears extremely short-lived in this restricted universe.

Validation: gross beauty, net catastrophe

The one-minute model was the strongest candidate:

The net equity approaches zero because the strategy repeatedly pays a cost hundreds of times larger than its expected spread. The exact compounded loss is less informative than the unit economics: every $1 of expected alpha requires roughly $327 of modeled execution cost.

Even the paper's lower four-basis-point cost example would be far above the break-even cost of this ETH–SOL translation.

Why can gross Sharpe be high when the strategy is useless?

Sharpe measures consistency relative to variability. It does not know whether an edge is large enough to cross the bid–ask spread.

With hundreds of thousands of observations, a tiny and stable average can look statistically impressive. High-frequency research magnifies this trap:

• sample size makes tiny effects significant;
• aggregation makes the gross curve look smooth;
• turnover converts a small fee into the dominant source of PnL;
• latency and capacity can only make the real implementation harder.

The correct order of questions is therefore:

1. Is the relationship statistically detectable?
2. Does it survive chronological OOS?
3. Is the conditional return larger than executable costs?
4. Is there enough capacity after market impact?

Our experiment passed part of question one and failed question three by two orders of magnitude.

Robustness: why Final stayed closed

• Train: 2021, 2022 H1, 2023 H1 and 2024 H1.
• Validation: 2025-08 through 2026-01.
• Final: 2026-02 through 2026-08—not loaded by the pipeline.
• Temporal independence: decisions were spaced by the holding horizon, so labels did not overlap.
• Causality: only the completed trailing BTC return entered each forecast.
• Cross-horizon check: the positive spillover sign existed only at one minute.
• Walk-forward / Monte Carlo: not used to promote the strategy because no horizon cleared the mandatory net-Validation gate.

Opening Final would not rescue the hypothesis. It would merely spend pristine data on a rule whose break-even cost is already physically implausible for our execution architecture.

What this result does—and does not—say

It does suggest that BTC can lead SOL at very short horizons. It also shows that statistical spillover is not automatically a trading edge.

It does not refute the published paper. The original study uses 30 cryptocurrencies, richer cross-sectional sorts, adaptive LASSO and PCA. A broad universe can create much more dispersion between top and bottom forecast portfolios than a single ETH–SOL pair.

Our rejected object is precise: the always-rebalanced, market-neutral, BTC-ranked ETH–SOL implementation under RLXBT's standard realistic costs.

The idea may deserve reopening under only three conditions:

1. a point-in-time universe broad enough to reproduce quintile sorting without survivorship bias;
2. verified maker/rebate execution with sub-basis-point all-in costs;
3. a new untouched period, without tuning the current Final sample.

Research trail

Tools called: get_ai_instructions → create_hypothesis → causal minute synchronization → clustered regression → cost audit → report_research_event → update_hypothesis → create_canvas_note.

IDEA MAP hypothesis: hyp_1786476063962_16 · Verdict: REJECTED · Final files opened: false.

Reproduce

Pipeline: research/paper_cross_crypto_lead_lag.py. Frozen result: research/results/paper_cross_crypto_lead_lag/summary.json. Source data: synchronized Binance-style one-minute OHLCV parquet files for BTCUSDT, ETHUSDT and SOLUSDT.

The lesson is not that high-frequency predictability is fake. It is more uncomfortable—and more valuable:

Sometimes the market gives you a real signal and no economically possible way to collect it.