Recurrent Neural Network (RNN)
A neural network designed for sequences that processes data one step at a time while carrying forward a hidden state, letting earlier inputs influence later outputs.
A recurrent neural network is an architecture built for ordered data such as time series, text, or price histories. Unlike a feed-forward network that treats each input independently, an RNN processes a sequence step by step and maintains a hidden state, a kind of running memory, that is updated at every step and passed to the next. This loop lets information from earlier in the sequence affect how later elements are interpreted, which is exactly what you want when the meaning of today's data depends on what came before it.
The structural insight is that the same set of weights is reused at every time step, so the network learns a general rule for updating its memory rather than a separate rule for each position. This makes RNNs naturally suited to variable-length sequences. In a market context, an RNN can in principle learn that a price move means something different depending on the trend and volatility that preceded it, capturing context that a static model would miss.
The classic weakness of a basic RNN is the vanishing-gradient problem: when sequences are long, the training signal that should connect distant past events to the current prediction shrinks toward zero, so the network struggles to learn long-range dependencies. This limitation drove the development of gated variants, most importantly the LSTM and the GRU, which add internal gates that control what to remember and what to forget, allowing useful information to persist across many time steps. In modern practice, when people say RNN for forecasting they usually mean one of these gated versions.
Hedgewing.ai uses gated recurrent architectures as two of the four models in its forecasting ensemble: an LSTM and a GRU, both descendants of the basic RNN. They are paired with non-recurrent approaches (a temporal convolutional network and an attention-based Transformer) so the ensemble does not depend on a single way of modeling sequence. All four read the same 45 engineered features and are validated through nightly walk-forward backtesting before their predictions are combined.
Related terms
Neural Network · LSTM (Long Short-Term Memory) · Gated Recurrent Unit (GRU) · Time-Series Forecasting
Back to the hedgewing.ai glossary · See AI stock forecasts