Capitalytics

Capitalytics collects several data series as part of its service, and builds linear regression models for the values that we track. Each linked monthly series shows the most recent six months of models. Independents are six months older than the dependent series: Y(t) = Σ a_i X_i(t−6). For a label month (today, Sept 2026), the test window of Y is the prior twelve months (Sept 2025–Aug 2026) and the train window of Y is the four years before that (Sept 2021–Aug 2025).

Series that are updated quarterly instead show four experiments over one year, stepping one quarter at a time (today, 3Q2026 through 4Q2025). For 3Q2026 the test window is the prior six quarters (1Q2025–2Q2026) and the train window is the five years before that (1Q2020–4Q2024). Dates on those pages are quarter-end observations.

Predictors are the other series on this list plus the natural log and square of those series — denoted by a “LN_” prefix and a superscripted "2" suffix — added one at a time to minimize training AIC, at most 5 terms (or n/8 if smaller), one transform per driver, and only when that term’s p-value is at most 0.1. A term is omitted when the magnitude of its coefficient is at most 1e-10, or when it is correlated above 0.8 with Y or with an already selected X.

Integrated series (S&P 500, Dow, Home Price Index, Commercial RE) are fit as ΔY. Quarterly series use one row per quarter. LN_ and squared predictors are clamped to their training range. Carried-forward extra samples are not treated as published actuals. Coefficients smaller than 0.001 are shown in scientific notation.

Each section is a collapsible block headed by the fitted equation, with stargazer output for the train window, MAE, RMSE, and R² impact on the test window, and a chart of historical actuals from January 2020, separate train and test fitted curves that are joined to the forecast at the window boundaries (the forecast continues from the last fitted point through the label month; OLS is shown when its test MAE beats or matches a last-print baseline). An intercept-only equation means no lagged series improved training AIC by at least 2 after the intercept (a mean change / random walk with drift for ΔY). Test MAE is omitted when the test window has only extra samples. (as of Sep 18, 2026)