Multi-Step Foreign Exchange Forecasting with a Transformer

Overview

This is a documentation-only public release accompanying an exploratory research project on autoregressive multi-step forecasting in foreign-exchange markets.

The system converts hourly market observations into a proprietary discrete representation and uses a decoder-only Transformer to forecast future market states.

The proprietary encoding procedure, vocabulary-construction rules, preprocessing pipeline, thresholds, mappings, model weights, and inference code are not included.

Public Documents

English

  • paper/forex_forecasting_summary_en.pdf

Italian

  • paper/forex_forecasting_summary_it.pdf

Main Experimental Result

The one-hour trading rule did not cover transaction costs.

The cumulative four-hour rule produced, on the already examined out-of-sample period:

  • 957 trades
  • EUR 13,055.20 gross P&L
  • EUR 9,276.41 transaction costs
  • EUR 3,794.39 net P&L
  • 491.58 net pips

Long and short positions both contributed positively.

This result is exploratory. The out-of-sample period has already been examined, and the result has not been confirmed on a new, fully untouched holdout period.

Scope of This Public Release

This repository contains descriptive documentation only. It does not include:

  • model weights;
  • vocabulary files;
  • tokenizer or encoder;
  • preprocessing logic;
  • datasets;
  • token mappings;
  • quantitative decoder mappings;
  • training or inference code.

The public material explains the research question, methodology at a high level, experimental design, economic results, and limitations.

Limitations

  • The reported out-of-sample period has already been examined.
  • Spread assumptions date from June 2022.
  • Overnight financing costs are excluded.
  • The simulation uses a fixed trade size.
  • Some euro conversions were unavailable; their pip results were retained.
  • The proprietary representation is intentionally undisclosed.
  • No claim of future profitability is made.

Intended Use

Research, discussion, and methodological review only.

This material is not financial advice and does not constitute an offer, solicitation, or guarantee of future performance.

Author

Alberto Altieri
Sisifo Capital

License

See LICENSE.

Citation

Altieri, A. (2026). Multi-Step Foreign Exchange Forecasting with a Transformer: An Accessible Research Summary. Sisifo Capital.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support