JobAI / Engineering notes
A trained model still has to earn its place
Separating numeric forecasts from sourced explanations, with four baselines as the test.
Research prototype · evaluation in progress
- Official data
- Forecast
- Evaluate
- Explain
The problem
JobAI targets Finnish quarterly job vacancies. Its first prototype is designed to forecast one, two and four quarters ahead. A convincing explanation cannot establish that a forecast is accurate; those are separate questions.
Keep the numbers traceable
Statistics Finland’s PxWeb data supplies the numeric inputs and target values. The workflow records the query, table metadata, retrieval date and response hash. This connects an experiment to the data snapshot it actually used.
Compare before claiming improvement
The evaluation plan compares fine-tuning with seasonal naive, last-value naive, linear regression and Prophet baselines. Rolling-origin evaluation respects time order. Training completion is a milestone; beating the baselines is a separate requirement that remains to be demonstrated.
Give explanations their own evidence
The planned retrieval layer uses approved official sources to explain the results. It does not replace the structured numeric data. This boundary keeps narrative generation from quietly becoming the source of the forecast itself.