Methodology
How MacroRadar's models work
MacroRadar publishes two related models: a regime classifier that describes the current state of the US economy, and a recession-probability model that estimates the chance of an NBER-dated recession ahead. This page documents where the data comes from, the discipline behind the numbers, how the models are validated, and — just as importantly — what they cannot do.
Overview
The two models share input data but serve different purposes. The regime classifier is descriptive — it labels the present state of the economy across multiple independent dimensions (growth, inflation, market, and financial conditions) and fuses those labels into a single regime. The recession-probability model is predictive — it estimates the probability of an NBER-dated recession over the months ahead.
Both are built from public economic data, validated against decades of history, and published with their full track record. Nothing about the inputs or the validation window is hidden: a reader with the same public data can check every classification we publish.
Data sources
Every input comes from a public, authoritative source. There are no proprietary data feeds and no opaque vendors — only primary economic data, pulled fresh from its originating institution.
| Source | What we use it for | Frequency |
|---|---|---|
| FRED (St. Louis Fed) | Rates, inflation, growth, labor | Daily / monthly |
| BLS | Inflation and employment | Monthly |
| BEA | Growth and price indices | Quarterly / monthly |
| US Treasury | Yields and real rates | Daily |
| CBOE | Market volatility | Daily |
| ICE / BofA | Credit spreads | Daily |
| NBER | Recession dating (reference labels) | Ad hoc |
The input universe is deliberately narrow. A broad indicator set is easy to assemble but hard to interpret; we favor a compact set of well-understood series that produces stable, explainable classifications across the full historical sample. The full source-by-source breakdown, with links and our redistribution policy, is on the data-sources page.
How the models work
The regime classifier evaluates each dimension — growth, inflation, market, and financial conditions — independently, using the indicators most relevant to that dimension. Each dimension produces both a label and a confidence reading. The dimension labels are then combined into a single regime through a transparent rule that a reader can inspect, rather than an opaque black box: if you can see the four dimension states, you can follow how the regime label was reached.
The recession-probability model is separate, with its own inputs and its own validation. It produces calibrated probabilities: a 30% reading means that in roughly 30% of historically similar conditions, a recession followed within the stated horizon. Because the model is built for interpretability, each published probability can be decomposed into the indicators driving it.
Both models are refreshed daily with the latest economic releases. No single indicator drives a classification on its own.
Point-in-time discipline
Economic data is revised. GDP, CPI, payrolls, and many other series are routinely restated months after first release. A model that quietly trains on revised history learns from numbers nobody actually had at the time — and looks far more prescient than it would have been in real life.
MacroRadar stores every observation with its vintage date — the date that value was first published. Historical classifications only ever see the data that was available at the time, never a future revision. This point-in-time discipline is what makes the published track record an honest account of what the model would have said in real time.
Validation
The models are validated against every NBER-dated recession since 1979. Validation is walk-forward: parameters are learned on an earlier period and then evaluated on later data the model never saw during fitting, so the out-of-sample record reflects genuine performance rather than curve-fitting.
We hold to two principles. First, no look-ahead bias — no information from the evaluation period influences any choice made during fitting. Second, publish everything — the full track record, including false positives and the lead time to each recession, is published rather than summarized into a single flattering number.
The complete per-recession breakdown lives on the recession-probability page.
What the models do not do
Every model has a clear scope and clear failure modes. Stating them plainly is part of the methodology, not a footnote to it. The fuller account lives on the limitations page.
The regime model is descriptive, not prescriptive
It tells you the current macroeconomic state. It does not tell you what to do about it. Historical asset-class returns shown by regime are base rates — what has happened in similar regimes before — not a forecast of what will happen this time.
Neither model anticipates exogenous shocks
The 2020 pandemic is the clearest example: no macro indicator anticipated it, so the models identified the contraction coincident with its impact on the data, not before. The same holds for wars, oil shocks, and other events whose cause sits outside the input universe.
Structural change can invalidate historical patterns
The models assume that relationships between indicators that held in the past continue to hold. That is a strong assumption. We treat genuinely unprecedented readings as low-confidence rather than absorbing them as a new normal, but we cannot fully control for structural change in the economy.
Recessions are rare
There have been only a handful of recessions in the historical sample. Any model trained on so few events has wide uncertainty around its estimates, and any out-of-sample test involves at most one or two new events. The track record is directionally informative, not precise.
The models are US-only
All inputs are US economic data. The regime and recession probability reflect the US economy. Assets with significant non-US exposure are partly driven by factors the models do not see. Broader geographic coverage is on the roadmap, not in production today.
Disclaimer
MacroRadar provides macro regime analysis. Nothing on the platform constitutes personalised investment advice, a recommendation to buy, sell, or hold any security, or a forecast of future market or economic events. The models are educational and analytical tools. Past model performance does not guarantee future performance, and historical returns by regime do not guarantee future returns.
Users are responsible for their own investment decisions. The cited data sources are public and authoritative, but MacroRadar makes no warranties as to their accuracy, completeness, or timeliness beyond republishing them as received.