wtf is finbench?

FinBench is a benchmark for evaluating synthetic financial time-series data. Instead of asking whether synthetic data looks superficially similar to real data, it tests whether the synthetic data preserves the statistical properties that matter for financial applications like risk modeling and strategy backtesting. The project exists because popular benchmarks for time-series generation reward models that produce smooth, safe-looking data. A model can score well on those benchmarks while failing to capture real market behavior such as fat tails, volatility clustering, and the leverage effect. These are exactly the patterns that drive financial outcomes, so missing them makes synthetic data unreliable for real-world use. FinBench replaces those generic checks with a suite of tasks grounded in established finance research. The benchmark runs seven tasks on a fixed panel of US equities and macro data with seven dimensions, 200 paths, and a 60-step horizon. Four tasks measure fidelity: whether the synthetic data reproduces the real data's distribution, stylized facts, distance metrics, and random-walk structure. Three tasks measure utility: options pricing accuracy, how well synthetic data preserves strategy performance rankings, and risk model backtesting through standard tests used in banking regulation. A model can do well on fidelity tasks and still fail on utility tasks, which is why the benchmark reports them separately. To submit a model, you generate synthetic returns for the FinBench panel, organize them into a specific directory structure using a provided template, run the scoring script, and open a pull request. You do not need to share your model's code or trained weights, only the synthetic outputs. The project maintains a leaderboard where Sablier, the company behind FinBench, lists three of its own models alongside external baselines. Each edition of FinBench is frozen at a git tag for reproducibility, and scores are tied to a specific version of the scoring library called finval. The code is released under the MIT license, with reference results under CC-BY 4.0.