Generate fictional seasons with Datasets and Synthetic Sports Replay

Datasets asks you to define a world rather than a single experiment: teams, a schedule, and the scoring model underneath. Nine sport adapters are available, each with an independent performance and pricing model — Poisson for low-scoring sports, Gaussian for high-scoring ones.

Choose the adapter that matches the question you are asking. A method that works on a Poisson soccer season and fails on a Gaussian basketball one has told you something specific, and that only means anything if the two generators genuinely differ, which here they do.

Generate the season and export it. Because you configured the process, you know the ground truth: the underlying strengths, the true probabilities, and the pricing assumptions. Any method you then run against the export can be scored on whether it recovered them.

This is the property real data cannot give you at any price. With a real season you can only ever check whether a method agrees with the outcome, which conflates being right with being lucky.

Synthetic Sports Replay steps through fixtures on a timeline. Because the generation is seeded, the same configuration replays identically — which is what makes it usable for teaching, for a written-up experiment, or for an archive someone else can reopen.

Reproducibility is the difference between a demonstration and an anecdote. If a result cannot be replayed by someone reading your write-up, it is a story about something that happened once.

The strongest use is adversarial. Generate a season, run your method, and check whether it recovered the parameters you set. Then generate a season where the method should fail — different scoring model, tighter margins, fewer games — and confirm that it does.

A method that never fails on synthetic data is not robust, it is untested. Finding the conditions under which it breaks is more valuable than another run where it worked.

Generated seasons pair naturally with the rest of the platform. Send selection rules at them in the Model Comparison Lab, check calibration on the Statistics pages, and use the discrepancy detector to see which signals your method is actually leaning on.

The whole platform shares one deterministic engine, so a seed carried between surfaces means the same world — which keeps a multi-step investigation coherent instead of comparing across three different random universes.

Everything Datasets produces is invented. The teams do not exist, the games did not happen, and the odds were never offered anywhere. When you publish, teach, or share an export, label it as synthetic — the value is in the known generating process, and that value evaporates the moment someone mistakes it for a record of real events.

Novus Odds is explicit that it is an informational research tool and does not claim real-world profitability. Exports inherit that framing.