LSTM / GRU

Best for: Sequence-heavy forecasting

How it works

Uses recurrent cells with gated state updates — an LSTM’s memory evolves as $c_t=f_t\odot c_{t-1}+i_t\odot\tilde{c}_t$, with forget $f_t$, input $i_t$, and output gates $o_t=\sigma(W[\cdot])$ controlling information flow. The gates let gradients propagate over long horizons, so the network captures long-range temporal dependencies that vanilla RNNs cannot. Trained end-to-end on sequence windows with MSE or quantile loss, LSTMs/GRUs forecast by unrolling the recurrence and reading a prediction off the hidden state at each step.

Common fields

Sensors · energy · finance