M∕L MACRO LAB

Hedge Fund Manager: Trial Run

Don't guess. Check the data.

Loading…

Change forecasts by modelWhat each model expects

unit: bp · positive = up

The bands show how far the models missed in the past (darker = where outcomes landed more often). Faded rows are comparison models; hollow rings are the same model computed another way.

bands = 30 / 50 / 70% quantile ranges of past forecast errors (darkest first). The share of past tests whose outcome fell inside the band — not a probability about the future · * = out-of-sample forecast fewer than 5 (no band) · faded rows = benchmark models · filled dot = minimum-error forecast, hollow ring = bias-corrected forecast

One-line summaryConclusion

Model registryWhich model got it right

Select a row to open its basis and tests

Model	Change forecast	Skill vs random walk	Skill vs long-run average	test p · verdictVerdict	First / second half	MAE	Directional accuracy	MZ slope	Out-of-sample forecasts

Contribution / bp

recursive out-of-sample forecastsPast forecasts vs actual

actual ━ / selected model ┄

expanding-window estimation after 72 months · only months whose outcome is known are used

Base model legend: Literature a model from the economics literature · Central bank a model from a Bank of Korea / Fed report · Data a regularity found by exhaustive search · Benchmark the comparison bar. Coefficients are re-estimated on this data.

Swap the two variables

Relationship between two variables · full period · no regime split

By regimeDoes the relationship change by regime?By regime

The same data, split by regime. The result above is the full period and does not change.

Scatter plotX and the later change in Y, one dot per monthScatter plot

Coefficients by sub-periodSame direction in every period?By period

Standardized coefficient (per 1σ) · ±2 standard errors

Cross-correlation by lag

Out-of-sample forecasts with X aloneForecasting with X aloneForecasting with this alone?

actual ━ / forecast ┄

Detailed figures

The rule library is an Expert-mode feature

Building your own models, saving hypotheses to test, and keeping a track record open in Expert mode. Expert mode is coming soon.

Rule library

Saving a hypothesis as a rule keeps testing it the same way. The period follows the horizon above; the test asks whether the relationship held over the full sample(forecasting power · causation). Price indices as % changes, rates as bp changes, other series quoted in % as pp — the same units as the Relations screen.

Select a rule

Coefficients by sub-period (±2 standard errors)

by horizon

Horizon	β (per 1σ)	t	Observations	R²

Supported = sign restriction · |t| ≥ 2 · sub-period sign agreement ≥ 75% · placebo p < 0.10, all met · Partly met = right sign but some criteria unmet (|t| ≥ 2: ‘significant · unstable’, else ‘sign only (not significant)’) · Wrong sign (not significant) = opposite sign but |t| < 2 · Significant, opposite sign = opposite sign with |t| ≥ 2. t uses non-overlapping observations. The placebo test shifts the cause by 24+ months, breaking only the timing, and counts how often a t this large appears.

Home ladder simulation

Returns Showdown