Generated Regime Evidence

This page is a view over generated artifacts, not a second hand-maintained set of research numbers. Provider snapshots feed a historical proxy classifier; that timeline feeds duration, transition, asset-return, exploratory ANOVA, and median-split stability calculations. Refreshing the canonical pipeline updates the values shown below.

What This Evidence Can and Cannot Establish

The generated timeline provides descriptive labels. It can answer questions such as:

• How many contiguous periods the proxy classifier produced
• How long those periods lasted
• Which state followed another state in the generated sequence
• How asset returns differed on dates assigned to each state
• Whether those descriptive patterns were similar on either side of the generated timeline's median date

It does not estimate classifier accuracy. There is no independent set of ground-truth regime labels in the repository, and the full-sample quantile boundaries use knowledge of the broad historical era. The temporal split is therefore a stability diagnostic, not an untouched trained-model test.

Historical Proxy Method

The canonical historical generator loads atomic FRED snapshots for CFNAI, CPI, the Federal Reserve balance sheet, VIX, and high-yield option-adjusted spreads. It applies documented conservative availability lags before forward-filling. The resulting history remains subject to an important limitation: current FRED revisions are used, not reconstructed ALFRED vintages.

The historical proxy recomputes equal-frequency quantile boundaries from the complete aligned factor sample. State counts, availability lags, interpolation, and the composite mapping remain documented method choices. It must not be silently conflated with the live product classifier, whose operational thresholds are not statistically calibrated. A vintage-data and live-calibration study remains separate future work.

FRED now limits its ICE BofA HY OAS response to three years. The generator retains the earlier daily credit factor from the prior generated artifact only before the first fresh availability-lagged observation and records the origin artifact hash, retained date range, current provider hash, and merge rule.

Generated Episode Durations

Each period is a contiguous run of one generated state. The table contains the median, interquartile range, period count, and total classified days directly from the duration artifact.

Durations describe this high-frequency proxy timeline. They are not forecasts of how long the current live regime will last.

Observed State Changes

A transition is counted only when consecutive classified dates have different states. Each row below reports an observed share among changes out of the stated source state. Overlapping uncertainty intervals are not claimed, and the shares are not calibrated probabilities for the current market.

Regime-Conditioned Asset Returns

For each asset and state, the backtest compounds only daily returns assigned to that state. It does not use the first and last price across recurring episodes, which would incorrectly include returns earned in intervening states. Annualized figures make samples comparable but can be unstable when a state has few days or episodes; the tables therefore publish both counts instead of qualitative sample badges.

Reflationary Expansion

Crisis/Liquidation

The dashboard exposes the same generated structure for every state. Historical returns are context, not a recommendation or expected return.

Exploratory ANOVA Diagnostic

The pipeline runs a one-way ANOVA on daily returns grouped by generated label and applies Benjamini–Hochberg correction across the asset-level omnibus tests. Pairwise inference is deliberately omitted because independent, identically distributed observations are not established.

Eta-squared is reported descriptively. A small p-value does not prove causality, forecastability, economic significance, or correct classification. Persistent labels, serial dependence, heteroskedasticity, and endogenous market inputs preclude a confirmatory interpretation.

Median-Split Temporal Stability

For each asset, the pipeline compares regime-level annualized returns across the two sides of the generated timeline's median classified date. “Regime-return correlation” is a correlation across state summaries, not a daily-return correlation and not classifier accuracy. “Same-sign share” is the fraction of comparable states whose annualized returns have the same sign in both eras.

These results can reveal instability worth investigating. They cannot turn the later era into an untouched holdout because the quantile boundaries are recomputed from the full aligned sample.

Reproducibility and Provenance

The canonical refresh:

1. Fetches atomic FRED and Twelve Data snapshots.
2. Records provider-response and normalized-file hashes.
3. Generates the historical proxy timeline.
4. Recomputes 16 canonical asset/regime files, duration and transition outputs, exploratory diagnostics, temporal stability, correlations, monthly backtests, and case-study observations.
5. Rejects non-finite JSON and incomplete artifacts before updating application data.

Generated artifacts record timestamps, input hashes, generator hashes, Git state, and package versions where applicable. This supports reproduction and change review; it does not by itself validate the economic model.

Principal Limitations

• No classifier-accuracy target: independent ground-truth labels are absent.
• Revised data: current FRED observations can differ from what was known in real time; conservative lags do not recreate vintage releases.
• Retrospective boundaries: historical and live distributional boundaries use their documented full samples. They are reproducibly generated, but no independent accuracy-calibration target exists.
• Endogenous labels: market inputs can mechanically relate labels to asset returns.
• Serial dependence and heteroskedasticity: daily observations do not satisfy simple iid assumptions.
• Multiple comparisons: the reported FDR correction covers the declared asset-level omnibus family, not every possible exploratory analysis.
• Annualization: short or sparse state samples can produce extreme annualized figures, which is why raw day and episode counts accompany them.
• Structural change: relationships can vary with policy frameworks, market composition, data revisions, and sample endpoints.

Interpretation Rule

Use the generated tables to describe the recorded sample, identify questions, and frame risk. Do not call observed transition shares predictions, temporal stability classifier accuracy, exploratory p-values proof, or annualized historical returns expected returns.

View the current dashboard →