How we test honestly.
The crypto research industry has a survivorship bias problem. Most platforms show what worked and quietly bury what didn't. Backtests are tuned until they look good. Our method is built to remove that bias — and to make our process auditable.
Six principles
Strict IS/OOS splits
Every hypothesis is split into in-sample (used to design and calibrate) and out-of-sample (frozen, touched exactly once). We verify there is no temporal leakage: the in-sample window always ends before the out-of-sample window begins.
A priori verdicts
Before we see any out-of-sample result, we write down the exact thresholds that will count as a pass, a fail, or ambiguous. The verdict is then applied mechanically. There is no room to move the goalposts after the numbers come in.
No parameter mining
One parameter choice, justified in advance — not a sweep of fifty until one looks good. If a strategy only works under a specific hand-picked threshold, that is overfitting, not edge.
Real costs
We model realistic transaction costs (default 0.40% round-trip: taker fees plus slippage) on every trade. Many published backtests look profitable only because they ignore what it actually costs to trade.
Predefined risk warnings
An automated detector flags known failure modes — low trade counts, concentration in a single symbol, holding-period drift, large IS→OOS degradation — before we interpret any result. The warnings are defined ahead of time, not invented to explain a bad outcome.
Public Strategy Graveyard
Every failed hypothesis is documented publicly with full results. Negative findings are a first-class output, not something to bury. This is the part of the method most of the industry skips.
Data
All market data is daily OHLCV from Binance spot markets, the deepest and most liquid venue in crypto. We use a single source deliberately: mixing exchanges introduces artificial steps in price series that can trigger false regime transitions and contaminate backtests.
- Universe~50 curated USDT pairs, ingested daily into a versioned PostgreSQL store. Real candles are flagged separately from forward-filled gaps.
- Survivorship biasDelisted assets are frozen at their final trading day — never deleted. Backtests see the market as it was, including the assets that died.
What we won't do
Our constraints are part of the product. They are what separate a research lab from a signals vendor.
- ✕Sell trading signals or alerts
- ✕Promise specific returns
- ✕Manage anyone's capital
- ✕Recommend specific buys or sells
- ✕Hide failed tests
Conflict-of-interest disclosure: Apex Quant does not hold positions on strategies it reviews, and does not accept paid placements in its research.