Same information · different forecasting methods
Follow a recorded monthly decision from its inputs to its prediction, risk adjustment and portfolio weights. Then compare what the models predicted with what actually happened.
Start here · choose a recorded decision
Changing the model or asset updates every view. Choose a month to inspect its exact calculation, or start with the largest forecast miss.
01 · what the model sees
An arrow marks an input clipped to the range learned from earlier training data. Other inputs pass through unchanged.
| Input | Training clip range | Used value in training standard deviations |
|---|
02 · why this forecast?
Left lowers this prediction. Right raises it. Values are contributions in percentage points of monthly return.
These terms reconstruct this fitted model's forecast. They are not causal effects. Displayed values are rounded.
03 · Black–Litterman adjustment
The adjustment uses all three asset forecasts, their past errors and how asset returns move together. It is not a single confidence slider.
04 · portfolio construction
Weights depend on the whole portfolio's expected returns, risk, trading penalty and bounds. A higher forecast for one asset does not automatically imply a larger position.
Stage 1 · essential comparison
Only the forecasting method changes. The historical mean uses the same completed return observations and ignores the predictors.
Click a month in this chart to inspect its decision above. The month picker provides the same control with a keyboard.
Out-of-sample R² across all assets. Above zero means lower squared errors than the historical mean; below zero means worse. Select a cell to inspect that model and asset.
| Core model | MAE | RMSE | OOS R² vs. mean | Forecast range | |Forecast| > 20% | Avg. active predictors |
|---|
MAE and RMSE are monthly return errors. OOS R² compares squared errors with the expanding historical mean on identical dates: positive is better, negative is worse. This uses a past-only mean forecast, not the realized test-sample mean. Active predictors exclude the intercept.
Stage 2 · change one thing
Lags and interactions are tested separately. The Ridge reference retains only the four additive core inputs.
Add the three asset returns known at the decision close, plus the previous month's growth and inflation surprises. Forward holding-period returns are never used as lags.
Add Growth surprise × ANFCI and Inflation surprise × ANFCI. This tests whether financial conditions change the relationship between surprises and returns.
No combined lag-and-interaction search or Bayesian VAR is run here. Review the focused comparisons before expanding the model.
| Ridge specification | MAE | RMSE | OOS R² vs. mean | Forecast range | |Forecast| > 20% | Avg. active predictors |
|---|
Same forecasts-to-weights rules
Shared covariance method, BL settings, bounds, monthly execution and costs. Each model's view uncertainty comes from its own earlier completed forecast errors.
Drawdown is the fall from each portfolio's earlier peak. Zero means a new peak; lower lines mean deeper losses. Line styles and visible models match the return chart above.
Each bar is one monthly allocation. Colors identify assets, not models. Select a bar to inspect that month.
| Model / reference | Net total return | Daily max drawdown | Avg. monthly turnover | Summed costs (bp) | Δ vs. mean portfolio |
|---|
Portfolio return differences are not risk-adjusted alpha. The fixed 45/35/20 and BL-prior-only references use the same dates and cost rules. Short episodes are not annualized; the full development history is the primary comparison.
Raw and BL returns are monthly forecasts. Historical error standard deviation uses earlier completed forecast errors. It is not a calibrated predictive interval. Numbers use SPY / IEF / LQD order.
| Holding month | Decision → entry → exit | Raw forecasts | BL estimates | Past error std. dev. | Weights | α / L1 ratio for selected asset | Net return |
|---|
Stability across periods
Calendar-year results come from the continuous run. Partial years are labeled and holdings are not reset at event boundaries.
| Year / months | Selected net return | Mean portfolio | Additive Ridge | Selected drawdown | Selected asset OOS R² |
|---|
Recorded before this comparison run
At least 36 completed training targets. Inner validation requires at least 24 completed targets at each validation decision. Scalers, imputation and clipping are refit inside each fold.
Ridge α: 0.01, 0.1, 1, 10 on standardized targets and mean squared-error loss. Elastic Net α: 0.01, 0.1, 1; L1 ratio: 0.2 or 0.8. Parameters are selected per asset using past validation errors.
Start 45/35/20; SPY 20–70%, IEF 15–60%, LQD 10–40%. Monthly covariance, BL τ 0.05 and risk aversion 2.5. Costs: 5 bp per dollar traded from drifted holdings.
Only holding months through December 2022. The January 3, 2023 exit quote closes December's position. It is not a January strategy evaluation.
Signal Research changes the information available to the existing model. Model Comparison holds a compact additive information set fixed and changes the estimator, then adds lags or interactions separately. Its feature set and shared clipping differ from the wider-input Ridge, so compare models within this tab rather than mixing numerical results from the two experiments.
The small one-lag Bayesian VAR is optional and not implemented here. It requires a separately specified small variable system and a comparison with Dynamic Ridge using matched information. SVAR, neural networks and broad tree searches are deferred.
Training targets are standardized separately per asset and converted back to decimal total returns before metrics and BL. Targets are never clipped. There is no supplied cash-return series. OOS R² uses the actual historical-mean forecast error as its denominator. These choices were recorded after earlier research was seen; this is exploratory evidence, not a frozen production model or an independent confirmation.