SM-LIIGA-GURU

The project's core idea is simple: instead of asking a general-purpose language model who will win a match, the prediction is built systematically from collected Liiga data and measurable variables.

The goal is to create a dedicated database, feature-engineering layer, prediction models and an evaluation system that allows each prediction to be assessed retrospectively.

A general-purpose AI can produce a convincing match analysis, but that does not guarantee that the prediction is based on correct information that was actually available before the match, or that its predictive quality can be measured objectively.

SM-LIIGA-GURU is designed around a prediction process in which the source data, model features and prediction timestamp can be traced.

Raw data → normalisation → PostgreSQL → feature engineering → prediction models → prediction snapshot → match result → retrospective evaluation

The system is intended to combine historical and up-to-date Liiga data in a relational database. A prediction may use only information that was available before the start of the match being predicted.

The SM-LIIGA-GURU data layer is based on a PostgreSQL relational database into which Liiga data is normalised and stored. The system builds pre-match features from available information, which are then used by prediction models to produce match-specific assessments.

One of the project's most important principles is temporal correctness: a historical model must not be trained or tested using information that became available only after the match being predicted. Model evaluation is therefore performed chronologically.

SM-LIIGA-GURU is in research and development. Current work focuses first on the data foundation, data sources, data model, feature engineering and validation methodology.

No claims about prediction accuracy or winning performance will be published until they can be demonstrated through controlled historical testing and later through genuinely forward-looking prediction tracking.

The next step is to expand data collection and the relational database, build temporally correct features and define the first baseline models. Only after that will more advanced models and AI-assisted contextual information be evaluated for their contribution to predictive performance.

Language models may later be used to interpret news, line-up changes and other text-based context, and to help explain predictions. An LLM is not, however, the system's primary data source or a replacement for the numerical prediction model.