Encoding Price Direction as BIP39 Tokens for Transformer Forecasting
Summary
The article describes a pipeline that turns price changes into binary sequences, groups the bits, maps them to BIP39 words, and feeds word indices to a transformer encoder. The stated aim is to predict subsequent market sequences and inspect word patterns associated with bullish or bearish conditions. It also outlines model components, data preparation, and training techniques such as normalization, gradient clipping, and learning-rate adjustment.
The author reports 73% accuracy on USD/JPY and describes recurring token patterns and dependencies across multiple candles. These claims are presented without enough detail here to assess the evaluation design, target definition, sample period, baselines, or out-of-sample robustness. The encoding also discards the size of price moves by retaining only whether each close rose or did not rise; chunk padding and tokenization choices may affect the resulting inputs. The approach is therefore an experimental representation and forecasting method, not evidence that market movements have a natural language or that the reported accuracy will generalize.
Key ideas
- Price changes are represented as up-or-not-up bits and grouped into fixed-length chunks.
- Each chunk is mapped to BIP39 vocabulary tokens for use as model inputs.
- A transformer encoder is proposed to learn token sequences and forecast later market behavior.
- The author reports 73% accuracy on USD/JPY, but the provided account omits evaluation details needed to judge robustness.
- The binary representation discards return magnitude and depends on chunking and token mapping choices.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.