We'll train a real LSTM neural network in your browser on 5 years of data. Watch it learn — then see why it fails.
R² looks great because stock prices are autocorrelated — today's price is always close to yesterday's. The model learned to copy yesterday's price, not predict the future. The proof: direction accuracy is ~50% — no better than a coin flip. It can't tell you if tomorrow is up or down.
When a stock drops far below its recent average (measured in standard deviations), the probability of a bounce is higher than random. Buy at -2σ, sell when it reverts to the mean. Repeat hundreds of times. Let the law of large numbers work.
Not one big prediction — just a small edge, repeated.
These are backtested results on historical data. Transaction costs, slippage, and regime changes can erode the edge in live trading. This is educational, not financial advice.