Real-Time Macro Regime Detection: Knowing Where You Are Without Borrowing From the Future

Georgi Kuzmanov Senior Equity Analyst & Founder Columbia MSFE · 13+ years equity research

Ask a macro trader what they would most like to know and they will usually name a number: next quarter’s GDP, the next CPI print, where the terminal rate lands. Ask what would actually change their positioning, and the honest answer is almost always coarser: is the environment expansionary or contracting, calm or stressed? Getting that one classification right — and getting it right on the day, not in the revised data three months later — matters more to a portfolio than any individual point forecast.

This is the third article in our Macro & Strategy series, and it builds directly on the point-in-time discipline of part two: a regime detector is worthless — actively dangerous, in fact — if it consumes data that had not been published yet. Regime models are revised-data magnets, because the series that describe the business cycle best (payrolls, industrial production) are exactly the series that get rewritten for months after each release.

So we will do this honestly. By the end you will have the exact rule our system uses — eleven macro series, each with its threshold, each read only as it was knowable on the evaluation date — the code that enforces it, the full 2016–2026 timeline of what the score actually said, and a measured answer to the question that justifies the whole exercise: does the same portfolio really behave differently across regimes?

Why Regime Beats Point Forecast

A point forecast has to clear a brutal bar: it must beat the consensus of thousands of professionals, then translate that edge into a position before the market reprices. A regime classification clears a much lower bar: it only has to describe the present accurately. It is nowcasting, not forecasting — and the information needed is mostly already public, sitting in market prices (yield curves, credit spreads, volatility) and slow official series (unemployment, sentiment, housing).

The payoff is not a trade signal. It is conditioning. Cross-sectional strategies — the rank-and-select machinery from our series opener — earn their returns unevenly: a defensive rank shines in stress and drags in melt-ups. Knowing the regime tells you which behavior to expect from the book you already hold, when a drawdown is regime-consistent versus a genuine warning, and (further along, when we discuss stress-testing) which historical windows are the fair comparison set. None of that requires predicting anything; it requires an honest label for today.

The Formula: A Regime Score From Point-in-Time Series

Our production regime rule is deliberately not a black box. It is a transparent weighted checklist — eleven indicators, each mapped to points by fixed thresholds, summed and normalized to a 0–1 score where 0 is deep risk-off and 1 is full risk-on:

regime_t = ( Σ_k points_k(x_k,t) ) / ( Σ_k max_points_k )

  where every input x_k,t satisfies:  available_from(x_k) <= t
  and k runs over the eleven series below (a series that has no
  available value at t simply drops out of both sums)

The eleven inputs and their actual thresholds, exactly as coded:

Notice the deliberate weighting: the four heavyweights (term spread, credit spread, VIX, and the second curve measure) are all Tier 1 market prices — observable in real time, never revised. The slower official series carry one point each. The composite leans on data that cannot lie about when it was known, and merely tilts on data that arrives late.

Determinism in Practice: Tiers, Vintages, and the Series We Refuse

The implementation lives in our data layer as Data.macro_regime(date), and its core is a helper that makes the availability rule physically unavoidable — the same pattern as the fundamentals as-of merge from part two, applied to macro vintages:

def macro_regime(self, date) -> float:
    """0 = deep risk-off, 1 = full risk-on.
    DETERMINISTIC: only data with available_from <= date.
    Tier 4 signals (ICSA, INDPRO, PAYEMS) are excluded entirely."""
    ...
    def _val(sid):
        """Most recent AVAILABLE value for series on or before date."""
        sub = macro[sid].loc[:ds]                 # observation <= date
        avail = sub[sub['available_from'] <= ds_str]   # AND published <= date
        return float(avail['value'].iloc[-1]) if len(avail) else None

    # 1  yield curve 10-2 (Tier 1: immediate)
    v = _val('T10Y2Y')
    if v is not None:
        mx += 2
        if v > 0.5:  score += 2
        elif v > 0:  score += 1
    ...
    return score / mx if mx else 0.5

Two lines of defense are visible here. The available_from <= date filter means each series is read as first published — the unemployment rate the model sees for March is the March print as released in early April, on and after its release date, never the value later revised into the database. And the docstring’s exclusion list is the second defense: Tier 4 series — weekly jobless claims, industrial production, nonfarm payrolls — are banned outright, not because they are uninformative but because their first prints are so heavily revised that even a vintage-aware model would be conditioning on noise. The tier assignments are driven from a config table of per-series availability windows, so the main series list is derived, not hand-curated per run.

The result is the property that justifies the word deterministic: for any historical date, the backtest’s regime score and a hypothetical live calculation on that date see byte-identical inputs. There is nothing to “refresh” retroactively; August 2022 scores 0.286 today, and it will still score 0.286 in five years.

What the Regime Score Saw, 2016–2026

We ran the production function at every month-end from January 2016 through June 2026 — 126 evaluations — via a committed evidence script. Labeling scores ≥ 0.7 risk-on and ≤ 0.4 risk-off: 65.9% of months were risk-on, 31.0% neutral, and just 3.2% risk-off. The mean score was 0.72. The decade, in other words, was mostly a bull backdrop punctuated by short storms — which is itself the first honest lesson: regimes are few. A classifier that fires “risk-off” four times in ten years gives you four data points to validate against, not four hundred.

The episodes are where it gets instructive:

• The minimum was not COVID. The score’s decade low, 0.286, came in August 2022 — inverted curve, wide credit spreads, hot CPI trend, sagging sentiment and housing: nearly every checklist item failing at once. The 2022 bear market was, to this ruler, the deepest macro deterioration of the decade; 2022 logged three risk-off and seven neutral months against just two risk-on.
• COVID never read as deep risk-off — and that is the honest cost of PIT. In March and April 2020 the score sat at 0.50: the Tier-1 market inputs (VIX above 25, blown-out credit spreads) collapsed instantly, but the slow official series — unemployment, sentiment, housing — had not yet published their catastrophe, and the Fed’s balance-sheet explosion was simultaneously adding points back. A model fed final revised data would paint March 2020 solid red; a model that only knew what was knowable read “severe stress, aggressive policy response” — which is what a live trader actually saw.
• 2021 was the cleanest signal of the decade: twelve risk-on months out of twelve. And 2023 shows the model’s conservatism: eleven neutral months while the index rallied hard — the curve was still inverted and CPI still hot, so the checklist refused to call it risk-on. A regime label is a description of macro conditions, not a market-timing oracle; 2023 is what the difference looks like.
• As of the June 2026 evaluation, the score stands at 0.786 — risk-on, consistent with the backdrop our current Week Ahead coverage describes from the narrative side.

Returns by Regime: The Same Book in Three Weathers

Conditioning is only useful if performance actually differs by regime. Our v7 validation layer answers this with a second, complementary regime lens: instead of macro fundamentals it tags each trading day bull / bear / high-vol from the market itself — realized volatility above 25% annualized is high-vol; otherwise price above or below its 200-day average decides bull versus bear. Both look-backs are rolling, so the tags are point-in-time by construction. We ran that code, unmodified, over our real ten-year low-volatility strategy backtest (the 20-name sleeve introduced in part one):

Each row compounds only the days carrying that tag (the slices are non-contiguous, rebased to a common start), which is why the numbers are large — they are cumulative over roughly a decade of days in each bucket. The pattern is the entire argument for regime awareness in two lines: in bear-tagged conditions the defensive sleeve lost half of what the index lost (−25.7% vs −50.7% cumulatively); in bull-tagged conditions it captured only about three-quarters of the index’s move. Same book, same rule, opposite relative behavior — the strategy’s character is a function of the regime, and no single full-period statistic can express that.

The Risk: Few Regimes, Wide Error Bars

Now the caveats, which for regime models are unusually sharp:

• Small samples, honestly small. Only 8.4% of days in a decade were bear-tagged — about 220 trading days, dominated by two episodes (2020, 2022). Any per-regime statistic is an estimate from a handful of effectively independent events, and its error bars are wide. Quantifying that uncertainty properly — bootstrap confidence intervals on regime-sliced metrics — is a later article in this series, but the caution applies now: treat per-regime numbers as direction, not precision.
• Threshold overfitting is one click away. Eleven indicators with hand-set thresholds is eleven opportunities to tune the past. Our defenses are structural — thresholds are round numbers with textbook rationales (an inverted curve, an 18-handle VIX), the weights lean on unrevisable market prices, and the same wall of validation gates that judges every strategy (multiple-testing discipline included) stands between any “improved” regime rule and production. The moment a regime model is tuned until the backtest looks best, it stops being a measurement and becomes the thing being overfit.
• The lag is the price of honesty. COVID reading 0.5 instead of 0.1 is not a bug to fix; it is what real-time knowledge looked like. Any “improvement” that makes March 2020 look better in backtest almost certainly works by borrowing information from April. Accepting the lag — and sizing the role of the regime signal accordingly — is part of the design.

The AlphaEdge Take

You do not need to out-forecast the market’s economists; you need an honest, deterministic label for the environment you are already in. Build it as a transparent checklist over point-in-time data — market prices weighted heaviest, revised series excluded — and it will tell you which behavior to expect from your book: our defensive sleeve lost half the index’s bear-regime damage while giving up a quarter of its bull-regime upside, and knowing that in advance is the difference between riding a drawdown and panicking out of it. Trust the label to describe, never to predict — and remember that its worst month, honestly computed, was August 2022, not March 2020, because honesty includes admitting what you could not yet know.

Georgi Kuzmanov

Senior Equity Analyst & Founder at AlphaEdge. Columbia University MSFE (2011–2013). Covering equities, macro, and geopolitics for serious investors.