Backtest

Anyone can claim a model works. This shows what it would have said before the outcome was known, using data the model never got to see in advance — hits and misses both.

Prefer a narrative read? See three of these as strategy memos →

The figures below exclude 32 calls from 2008-09 (global financial crisis), 2020 (COVID), and 2022 (energy-price/inflation shock) — years of a recognized global macro shock a simple trend model isn't built to predict. Every one of those calls is still shown, in full, in the lists further down, including as a miss.

159 of 160 total calls are tagged Backtest (a retrospective run of the model against real historical data, not a real-time call) — 1 are tagged Live, actually published predictions.

1 of 128 resolved calls predate the confidence-% field and aren't included in the calibration stats below.

Calibration

Raw accuracy can be misleading — a model that's right 70% of the time when it claims 95% confidence is a worse forecaster than one that's right 60% of the time when it claims 60%. Calibration checks whether stated confidence actually tracks with real outcomes.

All calls

Backtest = a retrospective run of the model against real historical data, done after the fact, using only data that would have been available at that point. Live = an actual call, published before the outcome was known. Never conflated.