We test so you don't have to · no. 1 (EN)

The edge exists. You just can't reach it.

We recorded Binance's order book for 34 days — 8.5 million rows. After eleven failed hunts for a trading edge, we finally found a signal that is statistically undeniable. It still loses 35 basis points per trade. This is the autopsy.

Over the past year we executed eleven candidate edges with pre-registered tests: momentum and breakout on every timeframe, grid bots from ads, funding harvesting, on-chain metrics, prediction markets with LLMs, mean reversion across 305 coins. The verdict was always the same — there was no signal, it only looked like one. Test number twelve came out differently. The signal is there. Real, stable, measurable. And that is exactly why this is the most valuable result of the whole series: it shows where the math ends and market structure begins.

§ 1

Why we recorded our own order book

The order book — the full ladder of resting buy and sell orders — is the deepest layer of information a market broadcasts. Academic literature rarely agrees on anything, but it agrees here: if anything predicts price on short horizons, it is the shape of the book. The catch: historical L2 data isn't free. Commercial archives run hundreds of dollars a month.

So we recorded it ourselves. One Docker container, a websocket on the full Binance book for BTC, ETH and SOL, one row per second: best bid/ask, spread, microprice, dollar depth within 5/10/25 basis points on both sides, book imbalance, trade flow. Over 34 days: 8.5 million rows — 429 MB, total cost zero.

§ 2

Rules are written before, not after

This is the most important paragraph of the article. If you search data and then pick what "works", you will always find profit — in the past. That trap ("in-sample charm") killed three of our candidates before we learned the lesson. Since then, an iron protocol:

• Pre-registration. Signals, thresholds, horizons and the success criterion are written down before the test runs. No tuning afterwards.
• Placebo control. We run 300 random twins against the strategy — same trade count, random times, random side. The strategy must beat the 95th percentile of chance, otherwise it measures luck, not information.
• Two time halves. Whatever doesn't hold in both halves of the data, on BTC and ETH alike, doesn't hold.

This time the rules were fixed: three book signals (depth imbalance, trade flow, microprice deviation), one uniform trigger — a z-score over 5 minutes, entry at ≥ 2σ — and three holding horizons: 10, 60 and 300 seconds.

§ 3

For the first time in eleven tests: the signal is real

The intuition behind the main signal is simple. When far more buy orders wait just below the price than sell orders above it, the price is more likely to move up in the next seconds — the book "knows" which direction is easier to walk through. In ten previous candidates, the placebo won. Not here:

Second view: the correlation of the signal with the subsequent move (Spearman IC). In quant research, 0.05 is considered respectable. Microprice at the 10-second horizon shows 0.24–0.33 — and, more importantly, nearly identically in both halves of the month. That is not an overfit model; that is a property of the market:

§ 4

And now the bad news

A basis point is one hundredth of a percent. The signal earns roughly one. What does a trade cost? Taker fees for entry and exit on spot come to 36 basis points, plus crossing the spread. Draw it:

The signal is thirty times stronger than chance — and thirty times weaker than the fees.

There is nothing to tune here. If the gap were 2×, you'd play with thresholds and smarter execution. But the gap is an order of magnitude: break-even needs costs below one basis point per round trip. The cheapest retail route — maker orders on futures — costs 4 bp, still four times more than the signal carries. The pre-registered gate said it plainly: NO-GO, passing = [].

§ 5

So who collects this edge?

Market makers. Firms with VIP fee tiers (zero to negative maker fees), colocated servers, and algorithms that hold queue position in the limit order book. For them, 1 bp per trade × millions of trades a day is a business. There is a deeper lesson in it: this edge is not a reward for clever prediction. It is a wage for a service — providing continuous liquidity, with all the infrastructure and risk that entails (a resting limit order gets filled preferentially exactly when informed flow runs against you — adverse selection). The market is not a random walk. It just charges for its predictability precisely what servicing it costs.

§ 6

What to take away

• Count costs before the signal. The average move your strategy captures must be a multiple of fees + spread. If it isn't, no entry tuning will save it — the disease is structural.
• Confront every strategy with a placebo. 300 random twins is a few dozen lines of code and the cheapest self-deception detector we know. Ten of our twelve tests were killed by exactly that.
• Write the rules first. A success criterion written after the test is not a criterion, it's an alibi.

Internally we call the standing conclusion "§7": on efficient large-cap markets, retail has no timing edge and the realistic goal is break-even. Twelve tests, twelve confirmations — this was the first that also showed why. The edge in the market exists. It just isn't collected by whoever predicts, but by whoever gets paid for a service.