Validation
How we know GridReplay is accurate
We test the oracle on six years of data, out of sample, and the engine behind it against real bots, order by order. We publish every result: the good, the bad and the ones we left out.
median order match by level and side across 33 real grids.
bots with a simulated net result within ±5% (or ±1 USDT) of the real one.
real liquidations reproduced, with the result within the margin.
filled orders compared one by one, across 33 long, short and neutral bots.
The oracle, six years out of sample
Does following the oracle improve results?
For each pair and each start (every 3 days since 2020), the oracle chose long, short or neutral from the positioning signals, recalibrated every month using only earlier data, exactly as it works on the site today. Each choice was simulated with this page’s engine for 30 days and compared with always running a neutral grid. Independent unit: the month, because crypto pairs move together.
per 30-day bot versus neutral; 95% confidence interval by month: +0.4 to +3.3.
months with enough data (2022–2026) in which the oracle beat neutral.
result of the worst 10% of bots, versus −24.1% with a fixed ±10% range (4,846 neutral grids).
of neutral grids with the balanced range, versus 55% with ±10%.
| Period | Months | Edge vs neutral | Data |
|---|---|---|---|
| Jan 2022 → Apr 2024 | 23 | +2.5 pp | never seen when designing the oracle |
| May 2024 → Jul 2025 | 15 | +1.1 pp | used to choose the signals |
| Aug 2025 → Oct 2026 | 14 | +1.3 pp | never seen when designing the oracle |
| Range profile | Width | Typical result | Profitable | Worst 10% |
|---|---|---|---|---|
| Aggressive | ±1,5σ | +2.7% | 68% | −15.4% |
| Balanced | ±2σ | +2.2% | 70% | −11.5% |
| Cautious | ±3σ | +1.6% | 71% | −6.6% |
| Fixed ±10% | — | +2.2% | 55% | −24.1% |
- Signals tested and dropped: funding, open interest, basis, DXY, price momentum and Fear & Greed showed no stable relationship. The ones kept (top-trader positioning and aggressive buying) passed the same test in all three periods.
- It’s an average edge, not a guarantee: 2024 was worse than neutral. 2x futures grids, Pionex fees and Binance 1-minute candles.
A real Pionex account, order by order
45 futures grids rebuilt from their history
With the trade history of a real Pionex account (March 2025 to October 2026) we can go beyond ROI: we rebuilt each grid from its orders and checked that GridReplay buys and sells at the same levels. The 12 bots that were modified while running are shown in grey and don’t count in the statistics.
| Pair | Direction | Year | Days | Orders | Match | Net difference | Status |
|---|
Other users’ bots
Results shared from the analyzer
Any Pionex user can run their history through the free analyzer and share their results anonymously. They show up here automatically, bad ones included. We can’t audit each submission the way we did with the account above, so we show them separately.
histories shared by analyzer users.
futures grids not modified while running.
median order match by level and side.
bots with a simulated net within ±5% (or ±1 USDT) of the real one.
| Pair | Direction | Year | Days | Orders | Match | Net difference | Status |
|---|
No submissions yet. Running futures grids on Pionex? Analyze your bots and be the first to add them.
Methodology and limits
What these tests show and what they don’t
- Real data: the trade history Pionex lets any user download. From it we rebuild each grid (range, grid count, mode, direction and order size) without using anything from the simulation.
- Prices: Binance perpetual 1-minute candles from the public archive, verified by checksum, with each contract’s real funding. Pionex aggregates prices from several markets, so small price differences explain part of the deviations.
- Order match: for each level and side, how many real fills the simulation reproduces, out of the total. Net result: grid profit plus position, minus fees, plus funding.
- We leave out 12 bots that were modified while running (investment added or range changed), because there is no single configuration to simulate. They appear in the table, in grey.
- It’s a single account: 33 bots over 19 months. It’s a serious test, but we keep extending it with bots from more users (through the analyzer) and want to add other exchanges.
- Validating the past doesn’t guarantee the future: GridReplay tells you what would have happened, not what will happen. None of this is financial advice.
Running futures grids on Pionex? Check GridReplay against your own bots with the free analyzer and, if you like, add your anonymous results to this page.
Spot something that doesn’t add up, or want more detail on a bot? Write to [email protected].
Test your own bot with this engine
The free simulator uses exactly this engine: it launches your setup across up to 100 periods from the past year and shows you how often it wins, how much it loses when things go wrong and how often it gets liquidated.
Open the simulator