PORTFOLIO LAB · HISTORICAL DATA
A range of futures.
A test against history.
Build a portfolio, forecast its return distribution, and see how the methods performed on earlier periods.
Portfolio outlook
Shaded area: model 10th–90th percentiles, not a guarantee. Solid line: portfolio median. Dashed line: simulated SPY median. Returns include the entered rebalancing costs, excluding initial and final trades.
Tested on earlier periods
| Method | Forecast error ↓ | 80% coverage | Loss Brier ↓ |
|---|
Forecast error is CRPS, which scores the full return distribution; lower is better. Coverage should be near 80%. Brier scores assess the probability of a loss; lower is better.
Current model weights and evaluation limits
Weights use only earlier forecasts whose full horizon had elapsed. Models refit at quarterly test dates. Paired block bootstrap intervals account for some serial dependence and overlapping outcomes. These are exploratory comparisons, not an untouched holdout or adjustments for repeated portfolio searches.
Testing today's chosen holdings introduces selection and survivorship bias. This is a forecast evaluation for fixed holdings, not a historical stock-picking strategy.
Data behind this run
| Macro state | Latest value | Observation | Vintage |
|---|
Employment growth and inflation are year-over-year percentages; unemployment is a percentage level. Philadelphia Fed vintage snapshots enter the model one full month after their vintage month. Market returns use Yahoo Finance split/dividend-adjusted closes as a total-return proxy; historical vendor corrections can revise the series.