Learning curve: risk-adjusted return on 40 unseen yearsSharpe ratio
One unseen year, played by the current agent-
Shading shows the hidden market mood: teal for a bull run, red for a bear market. The agent never sees it.
The market
Simulated, with Moods
Prices come from a generator with hidden moods: calm, bull and bear. Each mood lasts weeks or months, and bear markets are more volatile with occasional crashes. That gives trends and volatility clusters a real but imperfect edge, as in real markets, without pretending to forecast any real stock.
Q-learning
Learning by Trial and Error
The agent sees only the trend, recent momentum, volatility and its own position. For every situation it keeps a score for each allocation, the Q-value. After each day it compares what happened with what it expected and nudges the score toward the result. Early on it explores at random; it gradually trusts what it has learned.
The honest caveat
Why Test Years Matter
Results are always measured on 40 years the agent never trained on. An agent that only looks good on its training data has memorised noise. Real markets are far harder: patterns are weaker, change over time and get competed away once found.
Educational only. Every result on this page comes from simulated data. Nothing here is investment, trading or betting advice.