ForecastingStocks

Standard neural networks treat each input independently and have no notion of order. But a price history is a sequence — the order is the information. Recurrent neural networks (RNNs), and especially the LSTM (Long Short-Term Memory), are built to model sequences, and the LSTM is the model ForecastingStocks uses for its longer-horizon forecasts.

What makes an LSTM different

An LSTM processes a sequence one step at a time while maintaining a memory (a 'cell state') that carries information forward. Crucially, it has learnable gates that decide:

what new information to store,
what old information to forget,
what to output at each step.

This gating lets it remember relevant context over long spans — for instance, that volatility spiked twenty days ago — while discarding noise. Plain RNNs struggle with long memory because the learning signal fades over many steps (the 'vanishing gradient'); the LSTM's gates were invented precisely to fix that.

Why it suits price data

Markets have memory: volatility clustering, momentum, regime persistence. An LSTM can in principle learn these temporal dependencies directly from raw return sequences, without a human hand-coding indicators.

The honest caveat — even for our own model

The same dangers from the previous lesson apply with full force, plus one more: it is dangerously easy to leak future information into a sequence model through careless windowing or scaling, producing a backtest that cannot be reproduced live. ForecastingStocks treats its LSTM output as one probabilistic input among several, never a guarantee — and every forecast on the platform should be read that way. A model that fit the past is not a promise about the future.