Multi-horizon financial forecasting engine

The problem

In financial forecasting it is easy to produce a back-test that looks good and is wrong. Two things can make a model look better than it is.

It can see the future by accident during development, through data that would not have been available on the day. Or, when enough versions are tried, some of them look good by chance.

A team looking at a forecast, its own or a vendor's, needs to know which of these it is looking at before any money depends on it.

What we built

A full pipeline from raw financial time series to forecast to portfolio decision, built and tested by our team.

Shuffled-data controls. We rerun the same evaluation many times with the model's picks shuffled at random. A result counts only if it clearly beats those shuffled runs.

How it runs

The engine runs as a desktop application with its own database. It is written in Python.

Each walk-forward step retrains the models on data up to that point, forecasts the period after it, and moves on. The forecasts the engine is judged on are therefore always out of sample.

New ideas go through the same harness and are compared with our reference baseline before they are kept.

Results

We do not publish performance figures for this engine. We share them under NDA, together with the test setup that produced them, so you can check them.

Many of the improvement ideas we tested showed no real effect on these tests.

Nothing on this page is investment advice. We do not sell signals or forecasts to the public.

What we offer

We are not an authorised investment firm. We provide research tooling and evaluations, not investment advice.

Four ways to work with us, all under NDA. For any of them, you do not need to be a licensed institution.

Review of an existing forecasting model for overfitting

The smallest job we take. We put your current model, or its dated historical outputs, through walk-forward and shuffled-data tests and report whether its results can be told apart from chance. For the shuffled-data test the dated outputs are enough; you do not need to show us how the model works.

Shuffled-data (placebo) audit of your signals

For teams that only want the chance test: send dated forecasts (date, identifier, score). No code and no returns needed.

Research infrastructure

The pipeline itself, from raw time series to forecast to decision, with walk-forward training and the shuffled-data testing harness, set up on your data.

Forecasting models

Multi-horizon deep forecasting models built on your data, delivered with their full test record.

How a project starts

You send your model's dated historical outputs, or the dated forecasts alone, with a line on what the model is meant to predict. Within one working day we tell you which of the two fits, the review or the audit above. Either one is a first job of one to three weeks. Its scope and price are put in writing before work starts, and you pay nothing until you accept the written report. Infrastructure or model work is quoted later, when that report is in your hands.

What you receive

• A written report that says whether the results can be told apart from chance.
• The distribution of the shuffled runs, with your result set against it.
• A Sharpe ratio adjusted for multiple testing, for the number of trials you declare.
• How the signal decays across horizons.
• For a model review, the walk-forward result as well.

If you have not kept dated forecasts, start saving each one with its date, identifier and score, and we can say what else to keep.