Eight Edges, Three Live Feeds, Zero Alpha
I built a bot whose only job is to hunt for real trading edges — and to kill the fake ones before they ever touch money. I pointed it at real Binance data across eight edge families, three live data feeds, and three timeframes. It rejected every single one. This is the graveyard, with the numbers — because the honest negative is the whole point.
The machine, not the strategy
Most "trading bot" content sells you a strategy. I built the opposite: a continuous research agent whose default verdict is REJECT. An idea only survives if it clears a gauntlet — a real economic reason it should exist, a beat over a random-entry baseline, walk-forward out-of-sample validation, a parameter-plateau check, a 2× fee/slippage stress test, and a capital-scaling check. Anything that fails one gate is rejected and logged. The agent runs 24/7, refreshes its data, and when nothing passes it says so — literally: "No valid edge found yet. Continuing research." It never reports a profit it cannot defend.
What I tested — on real data, not a demo
Eight edge families across ten liquid coins, fed by three real live data feeds — perpetual funding (three years of history) plus open interest and the long/short account ratio (only about 30 days each, which turns out to matter — see below) — all pulled from Binance:
| Edge family | Verdict |
|---|---|
| Funding-extreme mean reversion | Rejected |
| Volatility-compression breakout | Rejected |
| VWAP mean reversion | Rejected |
| Cross-sectional momentum (long/short) | Rejected |
| Cash-and-carry / funding basis | Rejected* |
| Open-interest divergence | Rejected (data-limited) |
| Long/short-ratio crowding | Rejected (data-limited) |
| Cross-exchange spread | Monitor-only (not executable solo) |
The one that almost looked real (and why I still didn't trust it)
Time-series momentum — classic trend-following — is the kind of thing that looks great on a single backtest. So I walk-forwarded it: tune on the past, test on the unseen future, roll forward. Net of fees and funding, over six out-of-sample folds:
Worse, the "edge" flipped sign when I changed the lookback (+0.80% at 60 days, +0.44% at 100, −0.14% at 150). That is not an edge; that is a knob. And the alpha intercept vs. the basket was statistically zero while the beta was only −0.23 and unstable — so on ten coins that all move with Bitcoin, this was market exposure with a stop, not skill. Rejected.
"But lower turnover will save it" — it doesn't
The obvious objection to a losing strategy is that fees ate it: trade less and it works. So I re-ran every candle-based edge at 1-hour, 4-hour, and daily bars. Slower bars do cut the bleeding — but that is all they do:
| Edge | 1h | 4h | 1d |
|---|---|---|---|
| VWAP reversion | −25.5% | −14.5% | −5.8% |
| Funding-extreme | −15.6% | −10.5% | −3.6% |
| Vol-compression | −6.3% | +0.3% | 0.0% |
Every number gets less negative as turnover drops — and not one of them beats a random baseline or survives the cost-stress test at any timeframe. The lone tiny positive (vol-compression at 4h, +0.3% total) still fails the random-baseline and cost-stress gates — within noise, not an edge. The lesson is the headline: turnover was never the real problem. The signal was. Fewer bad trades is not a good trade.
I refuse to validate an edge on thin data
Two of the feeds — open interest and the long/short ratio — are only available for about 30 days of history from the free endpoint. A 30-day backtest is exactly how people fool themselves. So the agent has a hard rule: require at least 90 days across multiple regimes, or the verdict stays "data-limited — rejected", no matter how pretty the short window looks. Honest beats convenient.
What actually survived — and why it's not a salary
*The asterisk is the honest part. My agent's generic gauntlet — one fixed configuration applied to every edge — rejected its default cash-and-carry version too. But when I gave funding carry a dedicated, multi-year hold-and-collect (maker) walk-forward, it survived: holding spot, shorting the perp and collecting funding earned +2.0%/yr (taker) to +3.5%/yr (maker), 5 of 6 out-of-sample folds positive across three years. So it is the one real edge in the pile — a thin, market-neutral one, and strongly regime-dependent: annualized funding ran +7.2% (2023), +11.6% (2024 bull), +3.1% (2025), +0.5% (2026 bear). I am now forward paper-trading it live (maker); twelve days and 36 funding epochs in, it sits at about −0.01% cumulative — essentially flat in the current weak-funding regime, which is precisely what the backtest predicted for a bear stretch. It is real and low-risk — and at small capital it is coffee money, not income. I keep it because it is true, not because it pays.
Why a graveyard is worth more than a backtest
The rarest thing in retail trading isn't a winning strategy screenshot — those are infinite and mostly fiction. The rare thing is someone who built the machine to honestly kill bad ideas, ran it on real data, and will show you the bodies. Eight edges, three live feeds, three timeframes, net of costs, walk-forward — and a clean "no." That discipline is the entire job. It is also exactly what protects capital before a single euro is at risk.
If someone sells you a profitable crypto bot, ask for this: the walk-forward, across regimes, net of fees, and the list of everything they tested that didn't work. If they can't show you the graveyard, they haven't done the work.