The connectome experiment returned a clear negative. This is a separate line of work on the same data, following the one lead that came out of it — which turned out to be about execution mechanics, not biology.
Instead of predicting direction, wait for a real dip — price at least 2% below its recent high — then place a ladder of three resting buy orders below the market. Average whatever fills. Take profit at 1.2%, stop at 1.5%. Entries and take-profits are limit orders, which on this venue earn a small rebate rather than paying a fee.
Nothing about it is novel. It is mean reversion with patient execution. The question is only whether it survives honest testing.
| test | result |
|---|---|
| Walk-forward, parameters refitted each fold | +0.101% per trade · 49 of 60 folds positive |
| Every one of 12 markets, out-of-sample | all positive |
| Requiring price to trade through the level | +0.124% — unchanged |
| Strong downtrends | +0.077% — still positive |
| Fees tripled (0.15% taker + slippage) | +0.071% |
Five attacks, five survivals. At this point it looked like a real edge.
Every number above assumes a stop fills at the stop price. It does not — price runs past it before you are out. My first attempt modelled this as filling at the whole 5-minute bar's low, which turned the result negative on thin coins. But that was wrong: a stop fills within about a second of triggering, not across five minutes.
So I measured the real thing — how far price actually travels in the next second past a level, from public 1-second candle data:
| market | typical 1s slip | 99th-percentile |
|---|---|---|
| BTC | 0.0 bps | 3.3 bps |
| LINK | 0.9 bps | 8.3 bps |
| LIT | 1.3 bps | 10.7 bps |
| BEAT (thin) | 12.1 bps | 24.1 bps |
Applying each market's own measured slippage to the historical test:
| realistic slip | stress (p90) | positive | |
|---|---|---|---|
| liquid majors | +0.132% | +0.129% | 9 / 9 |
| thin alts | +0.051% | +0.046% | 2 / 3 |
It survives. Decisively on the nine liquid majors, marginally on two of three thin alts. Only BEAT stays negative, because its real per-second slippage really is 6–10× the majors'. The earlier "lethal" verdict was an artefact of measuring the stop over five minutes instead of one second.
The backtest assumed a resting order earns a rebate and a market exit pays 5.5 bps. Rather than guess, I pulled 10,000 real fills from 5 public accounts on Hyperliquid — a venue where every trade is on-chain and public — and measured what actually happened. (Addresses discarded on ingest; only anonymised behaviour analysed.)
Passive execution is real
One account rested 77.5% of its fills and came out with a net fee rebate. Sitting on limit orders and getting filled is not a backtest fantasy — a real trader sustains it. Across the five accounts the maker share ranged 25%–78%, which is a matter of how aggressively each chooses to cross, not of whether resting works.
Real costs were lower than I assumed
| backtest assumed | real (median) | |
|---|---|---|
| maker fill | −1.0 bps (rebate) | ≈ 0 bps |
| taker fill | +5.5 bps | ≈ +1.1 bps |
Costs were what killed the strategy on thin markets, and real costs are a fraction of the assumption — on this venue. So the fee side of the backtest was, if anything, too harsh.
And the liquidity split shows up in real fills too
By coin, liquid markets filled almost entirely passively at near-zero fees (HYPE 99.9%, ZEC 98.8%, AAVE 98.3%). The one illiquid name in the sample — LIT — filled only 4.8% passively and paid +4.5 bps to cross. That is exactly the pattern the backtest predicted: on thin markets you cannot rest, you cross and you pay. (Confounded by which account traded which coin, so: suggestive, not proof.)
This is not a validated strategy and nothing here is connected to an exchange.
300 days is one macro regime for crypto. The stop-slippage is now measured rather than guessed, but only over the few days of 1-second history the exchange serves publicly — a violent week could be worse than the sample. Entry fills still depend on order-queue position, which no candle or second-data can show; that needs live orders. And the structure is short volatility: small wins often, rare large losses — the shape that looks excellent right up until it does not.
What can be said, after two rounds of real execution data: a dip-ladder with maker execution shows a persistent edge on liquid crypto perpetuals — roughly +0.13% per trade out-of-sample, holding under measured second-level slippage and a p90 stress case, positive on all nine majors. On thin alts it is marginal to negative, driven by genuinely worse slippage. The liquidity split is the finding, and it is far larger than any signal the connectome produced. It is a lead worth a small live test on majors — not a validated strategy.
The per-trade edge held up under every test, so the honest next question is whether you can actually trade it. Three findings, and they do not all point the same way.
1. The dip timing is a real edge, not just limit-order luck
The sharpest control in this whole project: fire the identical ladder at random candles instead of on dips. Across all nine majors the dip filter beats random entry by +0.144% per trade — random is essentially zero. So it is the dip timing that carries the edge, not merely "limit orders on mean-reverting crypto".
| dip entry | random entry | |
|---|---|---|
| mean across 9 majors | +0.14% | ≈ 0.00% |
2. Walk-forward, every major is positive
Parameters refit on each rolling block and applied to the next unseen one: all nine majors positive, +0.03% to +0.11% per trade, 63–73% win rates. A single market compounds to about +16% over its out-of-sample period at 2% risk.
3. But as a portfolio it collapses — and the reason is the real lesson
Size those same winning trades into one book with realistic risk and a limit on concurrent positions, and it loses −79% with a 94% drawdown. That is not a contradiction. The nine markets' daily strategy returns correlate at +0.51:
| markets traded | 9 |
| correlation of their daily returns | +0.51 |
| effective independent bets | 1.8 |
Dips cluster. When Bitcoin drops, so do the rest — all your slots fill at once, right before a correlated leg down, and the losses stack. Nine "independent" edges are really one bet wearing nine hats, and the edge is too small to survive that concentration.
It does not, and that is the result. The bee's own signal was tested as an entry filter and made the strategy worse — adding it to the dip filter turned +0.064% into −0.022%. Triggering on the bee performed the same as triggering at random.
The connectome work was not wasted: it is what forced the controls, the walk-forward discipline and the habit of attacking a result until it breaks. Those are what found the liquidity split here. But the edge is in market microstructure, and the fly contributed nothing to it.