All articles · Guide
We froze our 5-minute robot for a week to check the backtest. Here is what we found
By Radiant Team, Radiant research ·
A backtest is only worth something if the live robot does the same thing. So before connecting real money to our 5-minute countertrend robot, we froze it for a week: no new settings, no new coins, no restarts. Then we ran the backtest on exactly the same days and compared it with the live model portfolio trade by trade.
What the comparison showed
Over the frozen days the live model portfolio closed about 3,400 positions and finished near zero. The backtest of the same days showed a clear profit. Entries matched: where both took a trade, the entry prices were practically identical. The difference sat in the exits.
We traced it to one detail of the backtest engine. On every bar it recalculated the take-profit level using the close of that same bar, and then checked whether the bar had reached that level. A live robot cannot do this: it places its take-profit using the level known at the previous close. The backtest was therefore taking quick profits that no real order could have taken.
What we changed
- We fixed the engine. Take-profits in the backtest now use only the level known before the bar, exactly like the live orders. With the fixed engine, the backtest of the frozen days matches the live robot.
- We recalculated and republished the year. The honest one-year result of the 5-minute robot on 50 coins is far lower than the first version we published: about +77 % on the model account instead of +511 %, with 10 of 13 months positive instead of 13 of 13, and a deeper maximum drawdown. The robot page now shows these numbers and says what changed.
- We added two protections to the live robot, chosen on the honest backtest and checked against the October 2025 crash:
- Portfolio balance. When the whole portfolio leans too far to one side, new entries on that side wait until the book rebalances. One-sided books were behind most of the large drawdowns we studied.
- Leverage-crowded coins. When a coin's open interest is far above its own monthly average, the robot opens no new shorts on it: shorts against a crowded, leveraged move did badly in our data, while countertrend longs there did well.
What this means
The robot's edge is real but much thinner than the first backtest suggested. We are not connecting real money until the live robot, with the new protections, matches the honest backtest for long enough. Everything is visible as it happens: the live model portfolio and every trade are public.
We would rather publish a smaller honest number than a big wrong one. Checking the backtest against live trading is now a permanent part of how we work, for every robot and every market we add next.
Figures are model portfolios and backtests, not results of client accounts. Past results do not guarantee future results.