GridReplay report · October 2026
We put Pionex’s AI agents to the test
Their assistant Pai suggests grid bot configurations and their agent Quant promises backtests. We replayed their suggestions over real market history, minute by minute, to see what would have happened to the money.
NEAR/USDT, last 30 days, 2-hour candles. The horizontal lines are the 36 levels of the geometric grid Pai suggests (3.60–6.00 USDT, 35 grids). Every time the price crosses one, the bot buys or sells.
Three takeaways
of months end in profit when Pai’s method is applied again and again over two years. On average, it loses money: −2.2% a month on SOL and −1.8% on NEAR.
of months are profitable on SOL when entering in the lower third of the range, exactly what Pai presents as the favorable moment. All other entries: 64%.
median match between GridReplay and 33 real Pionex grids, order by order. Quant, by its own account, doesn’t simulate funding or liquidation and has never been compared with real bots.
1 · What Pai suggests
Well-written technical analysis, no simulation
We asked Pai, the Pionex.AI assistant, for assets suited to grid trading and a specific configuration. It picked SOL and NEAR for their volatility and proposed ranges based on support, resistance, RSI and Bollinger Bands. Its market data was correct: the lows it cited (92.35 and 3.91 USDT) had been set that same day.
| Pai’s suggestion | Range (USDT) | Grids | Mode | Stop loss |
|---|---|---|---|---|
| SOL/USDT spot | 85 – 125 | 45 – 65 | geometric | 80 |
| NEAR/USDT spot | 3.60 – 6.00 | 35 – 50 | geometric | 3.30 |
What it doesn’t give is any result figure: not how much that configuration would have made or lost, not how often the stop gets hit, not how it compares with simply buying and holding. So we calculated it. First, its exact configurations over the last month:
| Last 30 days | Result | From grid | From position | Buy and hold | Cycles |
|---|
This test favors Pai: its range comes from that very month, so the price never left it. Even so, the grid profits from the swings while the inventory it bought loses value as the price falls. It beats buy and hold, but doesn’t make money on SOL.
2 · The method, over two years
We repeated its reasoning 200 times
A single configuration says little: what matters is whether the approach works repeatedly. We turned Pai’s reasoning into a fixed rule, derived from its two suggestions, and applied it every week since late 2024, holding each bot for 30 days:
- Lower bound: 8% below the low of the previous 30 days.
- Upper bound: the high of those 30 days.
- Geometric grid with ~1% between levels, stop loss 7% below the lower bound, 0.05% fee.
30-day result of each bot, as % of investment
| Asset and entry | Bots | Profitable | Median | Mean | Worst | Stop hit |
|---|
The outcome looks like a coin toss: a little over half the months end in profit, but the losses are bigger than the gains and the average is negative. One bot in four hits its stop loss, losing around 20%.
And the advice to enter when the price is in the lower third of the range makes results worse on SOL and doesn’t improve them on NEAR. When the price is near the monthly low, it’s often because it keeps falling.
3 · What Quant says about itself
AI-written backtests, never validated
Quant is Pionex.AI’s paid agent for strategy research. We asked it for the same NEAR backtest and it replied that backtest execution was disabled, without making up numbers, which is to its credit. Then we asked how it works. Here is a summary of its answers (quotes translated from Spanish):
- For each request it writes a new Python script, which Pionex’s engine runs candle by candle.
- It describes this as
a Python model of the grid idea
, not as the real bot. - It uses a single fee per trade, without distinguishing limit orders from market orders, and has no evidence that it calculates funding or liquidations.
- On its accuracy versus real bots:
I don’t have any validation figures
. - It has no access to the user’s bot history.
| Pai | Quant | GridReplay | |
|---|---|---|---|
| Simulates results | No | Yes* | Yes |
| Maker and taker fees | — | No | Yes |
| Futures funding | — | No | Yes |
| Liquidation | — | No | Yes |
| Validated against real bots | No | No | Yes |
| Analyzes your bot history | No | No | Yes |
4 · Is our simulator accurate?
Checked against 33 real bots
Asking for rigor means applying it to ourselves. We took the trade history of a real Pionex account (March 2025 to October 2026), rebuilt every futures grid from its orders and replayed it in GridReplay on Binance 1-minute candles, with the actual funding.
45 grids rebuilt. We left out 12 that were modified while running (investment added or range changed) because they aren’t comparable. That leaves 33: 18 long, 13 short and 2 neutral, on BTC, LINK, AVAX, HBAR, WLFI, XRP and others, with 2,298 filled orders.
Difference between simulated and real net result, per bot
median order match by level and side between simulation and reality.
bots with a net result within ±5% (or ±1 USDT) of the real one.
liquidations reproduced, with the result within the margin.
Methodology and limits
What we measured and what we didn’t
- Data: Binance 1-minute candles (spot for SOL and NEAR; perpetuals and funding for the validation), downloaded from its public archive and verified by checksum.
- Within each minute we assume a conservative path (on bullish candles, low first and then high). Orders fill when the price touches their level.
- The rule for Pai’s method is derived from two examples and was tested on two assets. Bots from consecutive weeks overlap, so there are about 25 truly independent periods per asset. It’s a clear signal, not definitive proof.
- Pionex aggregates prices from several markets; small differences with Binance explain part of the deviations in the validation.
- Past performance does not guarantee future results. This report is not financial advice or an investment recommendation.
Check your own configuration
GridReplay replays any grid over real market history, minute by minute, and analyzes the bots you’ve already run from your Pionex history. It’s in development: join the list and you’ll use it for free during the beta.
Join the waitlist