Long Short-Term Memory
LSTM is a gated RNN whose cell state can carry information across many steps, with input, forget, and output gates trained by gradient descent.
Definition
LSTM (Hochreiter & Schmidhuber, 1997) adds a linear cell state and multiplicative gates so gradients can flow through time without immediate vanishing. GRU later simplified the gate set. Both were the default sequence models until Transformers.
Why it matters
If you read any 2015–2017 NLP or speech paper, the backbone is LSTM. In finance, LSTM still appears in execution and vol papers because it is smaller than a Transformer on short series — not because it is more principled.
Case
TIMIT and machine-translation results jumped when LSTMs replaced vanilla RNNs. The same architecture, dropped on a 20-year daily equity panel with 30 names, will still overfit if you ignore sample size.
How to read it
Count effective sequence length and parameters versus N. An LSTM is not “deep learning alpha.” It is a sequence inductive bias that still needs a honest test.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Long Short-Term MemoryHochreiter, Schmidhuber · 1997 · Neural ComputationThe original LSTM paper (PDF).