The demo our AI failed
By Robert Yenokyan,
Our AI agents failed miserably in a client demo.
The client was disappointed in our AI.
It turned out the client's result was the wrong one.
The client had a trading strategy with good results. It was written up and coded. They wanted to see whether our agents could rebuild it and get the same result. And how long that would take.
Our agents rebuilt it. Their result was nowhere near the client's.
So we dug into the client's implementation and found a couple of small things. The kind that look like no problem at all.
Their strategy looked at the closing price of each candle, the bars on a price chart, to decide when to trade. Then it got in and out at that same closing price.
But you only know a candle's close once the candle has finished. After that, the decision takes a moment to compute and the order takes a moment to reach the exchange. That delay is called latency. Their test had none, so it filled every trade at a price you could not count on getting.
On one trade the difference is tiny. Over millions of trades it adds up.
A proper test also charges realistic fees. And it counts the cost of your own order pushing the price against you as it fills. That cost is called market impact.
With the delay, realistic fees and market impact counted across all those trades, the strategy was not worth trading.
A tool that copies your result copies your mistakes along with it.
Our agents are built to call our validated backtesting libraries instead of writing that code themselves, wherever the library has what they need. Those libraries account for the delay, realistic fees and market impact. That makes the agents less prone to wrong results than an AI writing everything from scratch.
The agents are our AI copilot, and it is in private beta. They design, run and check trading research on our infrastructure. You see what they built and what came out in a web portal, or call them from your own code through an API.
Founding users help shape the product and their requests get priority. We talk to everyone who applies.
If you have a strategy with good results and want to know whether they survive a rebuild, email robert_yenokyan@bonton.ai with what you would use the copilot for.
You can check your own backtest for this without any tools. Take a few trades and compare the price each one was filled at with the price its signal was computed from. If they are the same number, the test is filling trades at a price you could not count on getting.