Skip to content

FinIR 0.1.0: a compiler target for AI financial computation

Latest
Latest

Choose a tag to compare

View all tags
@Lethabo-Scofield Lethabo-Scofield released this 01 Sep 22:52
· 13 commits to main since this release
v0.1.0
299d2e2

FinIR 0.1.0: a compiler target for AI financial computation

First public release of FinIR — a financial intermediate representation and
incremental execution runtime for AI systems.

pip install finir

What's in it

  • Finance-typed IR. A typed computation graph with a .finir text form and a
    lossless JSON interchange format. The type system enforces the algebra of finance
    at compile time (money - money → money same-currency; money / money → ratio;
    money + days and USD + ZAR are errors).
  • Incremental dirty-set runtime. Changing one input invalidates only its
    downstream cone; the next evaluation recomputes exactly those nodes and reuses the
    rest in O(1). On the reference machine this is 1.7×–2.2× faster than full
    recompute with up to 99.6% cache reuse (measured; see docs/performance.md).
  • Dependency-aware reuse & a finance-native cache with hit/reuse metrics.
  • Scenario execution. what_if, named scenarios, and vectorized
    run_scenarios over million-row batches (~1,000,000 scenarios in ~46 ms on CPU).
  • Structured intent contract (schema v1.0). A canonical, versioned JSON envelope
    that a natural-language layer emits and the runtime validates and executes
    (apply_intent), with valid/ambiguous/unsupported/invalid statuses so
    vague language never becomes invented numbers.
  • CPU backend (NumPy) — the default, needs nothing else.
  • Optional GPU backend (CuPy) behind pip install "finir[gpu]", with a
    workload-aware dispatch planner.
  • CLI (finir run/compile/inspect/graph/benchmark/doctor).
  • Kernel library (arithmetic, corporate finance, working capital, time value of
    money, basic risk) plus a @finir.kernel extension point.
  • Benchmark suite and two research experiments (incremental reasoning;
    backend dispatch) — all numbers measured, never hard-coded — plus a critical
    prior-art analysis.
  • Apache-2.0, Python 3.11–3.13, CPU-first.

Honest caveats

  • GPU performance has not yet been verified on CUDA hardware. The GPU backend is
    optional and unit-tested via guards; its dispatch threshold is a heuristic pending
    measurement (see research/experiment_002_backend_dispatch.md).
  • FinIR is deliberately not a quant library; the kernel set is small.
  • Positioning vs. spreadsheets, incremental-computation systems, JAX/XLA/MLIR,
    QuantLib, and planning engines is examined in research/prior_art.md. We make no
    claim of being a first or a breakthrough — the novelty is a compositional
    hypothesis pending a formal prior-art review.

Links

Assets 4
Loading