Skip to main contentThe Local Bench

Use case 13 of 13

Personal and small-business finance analysis

Categorise bank exports, ask questions in plain English and forecast cash flow, with your statements never leaving your machine.

A calculator, a notebook with a pen and a black mini PC on grey paper
Image: generated with Grok Imagine

The real example

Actual Budget exports + Chronos-2 forecasting

Actual is an open-source, local-first personal finance app that imports bank files. Once transactions are on disk, a local language model can categorise them and answer questions, and a time-series foundation model can forecast. AutoGluon's tutorial shows Chronos-2 forecasting zero-shot, without training on your data, and fine-tuning when you want more accuracy. On GIFT-Eval, our calculation from the published results files puts Chronos-2 at 0.698 MASE relative to a seasonal-naive baseline (lower is better), with TimesFM 2.5 at 0.705.

Sources: Actual Budget on GitHub · Actual Budget: importing transactions · AutoGluon: forecasting with Chronos-2 · GIFT-Eval results

Why run it locally

Bank statements are among the most sensitive data you have. Forecasting models are tiny and run on a CPU.

Recommended stack

Three tiers built from our builds. Weights-in-memory is our arithmetic (parameters × bits), checked against each device's usable memory; it excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Budget

Silent home server

$1,283Sample

16 GB on GeForce RTX 4060 Ti 16 GB

Weights ≈ 5.3 GB of 15.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Chronos-2 is about 120M parameters; the LLM does categorisation.

Runtime: Actual Budget + AutoGluon + Ollama

Mid

Used-3090 24 GB

$1,980Sample

24 GB on GeForce RTX 3090 24 GB (used)

Weights ≈ 18.7 GB of 23.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

A coder model can write the pandas analysis for you; you check it.

Runtime: Jupyter + AutoGluon + llama.cpp

Pro

Mac Studio M5 Max

$2,499Sample

36 GB on Mac Studio M5 Max (base)

Weights ≈ 20.5 GB of 27.0 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Compare two forecasters and keep the better one for your series.

Runtime: Jupyter + AutoGluon

Setup outline

  1. Export transactions (CSV or OFX) from your bank or from Actual.
  2. Have a local LLM propose categories, then review and lock the rules.
  3. Aggregate to weekly or monthly series and forecast with Chronos-2 via AutoGluon.
  4. Ask the LLM to explain variances, but compute the numbers in code, not in the model.

Limits

  • Not financial advice. Forecasts are statistical, and one-off events break them.
  • Language models are unreliable at arithmetic. Always compute totals in code.
  • GIFT-Eval aggregates many domains. Your household cash flow is one short, noisy series.

Next: A private voice assistant for the home

Build prices come from our build datasheets: GPUs and devices are sourced retail or sold prices; other parts are sample prices until sourced, and the flag on each total says which. Model links go to our datasheets, where each parameter count and licence was checked against the Hugging Face API.