Forecasting portfolio
Forecasts you can explain, measured in dollars
A series of treasury forecasting projects, each built from scratch on synthetic and public data. Every project starts simple, adds one fix at a time, and measures what each step is worth with a walk-forward backtest and a dollar metric.
Projects
Project 1 · Seasonal baseline
Weekday-seasonal forecaster
Daily payment volume, 1–30 days ahead, with one explainable formula: recent level × weekday effect × payday effect.
−35% errorvs repeating last week · buffer $15.8M vs $26.5M a day
Project 2 · Spikes and holidays
Cash-flow forecaster
Builds on Project 1 in five steps: spike detection, holiday effects and Fourier seasonality with Prophet, across three corridors.
−38% errorvs the Project 1 baseline · $10.5M less cash held a day
Project 3 · In progress
FX margin forecaster
A two-stage model for FX margin and margin at risk: trend first, then gradient boosting on what the trend misses.
How each project is built
- Simulate data with known structure, so each model component can be checked against the truth, and run the same pipeline on public U.S. Treasury data.
- Start from a simple, explainable model and compare it with naive baselines.
- Add one fix at a time, each aimed at a specific kind of day the previous version gets wrong.
- Measure with a walk-forward backtest: refit at every origin, never use data from the future, and report errors by type of day.
- Translate into dollars: the cash buffer needed to cover shortfalls, and the cost of idle cash.