Research / FPL data
Ask the FPL database anything
Natural language meets live data, executable queries, and visible limits
FPL data is public, current, and difficult to question well. A simple request can require several tables, a historical snapshot, joins, filters, and a chart. SportSQL lets a reader ask in plain English, then translates the question into executable database work.
The system can answer many simple questions accurately. Its benchmark also shows exactly where natural-language sports analysis begins to break.
From a question to evidence
SportSQL ingests the public FPL API into a time-indexed relational database. It resolves player and team names, generates SQL from the question and schema, executes the query, and chooses a visual form when the result needs one.
The language model does not answer from memory. The database produces the result. This makes the path inspectable and reduces the risk of confident invented statistics.
- 01Resolve names and intent
- 02Generate executable SQL
- 03Run against current FPL data
- 04Return a table or chart
A benchmark built from live data
The authors introduce DSQABench with more than 1,700 questions. Each includes a gold SQL program, answer, and database snapshot. That last element matters because sports facts change. An answer can be correct for one date and wrong a week later.
The best reported system reaches up to 80% exact-match accuracy and a macro F1 score of 0.75 across its structured question-answering evaluation.
A live sports answer needs a date as much as it needs a number.
00.football interpretation
Complexity exposes the weakness
Single-step questions reach 93% accuracy. Two reasoning operations reduce that to 67%. Beyond three, performance settles near 50%. Queries involving table manipulation fall as low as 15.4%.
The pattern is more useful than one headline score. Retrieval and sorting are dependable. Aggregation, joins, and several linked conditions remain fragile. The interface should show the generated query and make complex answers easier to audit.
A research interface for FPL
The paper points toward a useful form for 00.football: articles that can be questioned. A reader could inspect a finding, change the player or gameweek, and see the query behind the new result.
The system is not accurate enough to hide the machinery. Its value comes from joining natural language with verifiable structured work, then showing where that work remains uncertain.
Strong on simple questions
This 2025 preprint evaluates a prototype centered on English-language Premier League data. Accuracy declines sharply for joins and multi-step calculations. Live schemas, transfers, and context limits can also change the result. SportSQL is best understood as an inspectable research interface, not a fully reliable autonomous analyst.
Read the original research ↗