Agent 02 · Online neural network
The market
Each day has 390 one-minute bars. Prices wobble around a fair value and drift back to it (mean reversion). News occasionally moves the fair value, and the price takes several minutes to catch up (momentum). Every change of position pays the cost you set.
The network
A small neural network (about 100 numbers) reads the last 1, 5 and 15 minutes of returns, the distance from the day's average price and the time, then forecasts the next 10 minutes. It goes long or short only when the forecast is big enough to pay the cost. Nobody tells it the rules.
Learning from errors
Every minute of every practice day it makes a forecast, waits ten minutes, and measures how wrong it was. Each weight moves a little in the direction that would have shrunk that error. Most minutes are noise, so it learns slowly, and forecast skill stays small, as it does in real markets.