Timing Signal Composites with Macro Regime Predictions

This article extends our previous analysis from “A Practical Approach to Weighting Signals”¹ by introducing a forward-looking macro regime layer on top of it. Our prior work began from a deliberately narrow selection: 12 signals, 4 apiece within Sentiment, Quality, and Value composites, bound by explicit signal- and composite-level constraints and set against one another using the FactSet Optimal Weights Engine (“OWE”) under 3 competing philosophies of Equal Weight, Risk Parity, and Maximum Information Ratio (“Max IR”). It treated the investable universe as facing a single, unconditional set of expected returns and risks through time.

This follow-up study asks a related but distinct question: Can a forward-looking macro regime signal be used to time the composites themselves so that the weighting process adapts to whether the economy is expected to be expanding or contracting?

The premise draws on a well-documented, if contested, strand of factor timing research. Hodges, Hogan, Peterson, and Ang² found that combining several regime and valuation predictors is more effective at timing smart beta factors than relying on any single indicator. Practitioner work from Newfound Research³ and MSCI⁴ has similarly argued that momentum and market-sensitive factors tend to lead in expansions while quality and defensive factors hold up better as growth slows.

Sheth and Lim⁵ add a note of caution from the academic side: While Fama-French factor returns differ meaningfully across the business-cycle stages they define, the term spread they use to predict those stages showed no statistically significant relationship to the NBER recession indicator in their sample. It underscores that identifying exploitable factor patterns after the fact is easier than forecasting the regime that produces them in advance.

This study does not attempt to resolve that debate. Instead, it documents what happens when a specific, disciplined regime prediction process is layered onto the same OWE framework used previously, so that the reader can judge the trade-offs directly.

The Regime Signal

The regime framework is deliberately simple: two states, Expansion and Contraction, defined by the ISM Manufacturing PMI relative to the 50 threshold, smoothed with a three-month rolling average and shifted one month to reflect publication lag. Over the full 1948 to 2026 history, Expansion has accounted for roughly 68.5% of months and Contraction the remaining 31.5%, with a lag-1 persistence rate of 93.6%, meaning the regime rarely flips from one month to the next.

A CPI-based regime framework was considered and rejected. In separate prior work not reproduced here, inflation regimes proved approximately 97.9% persistent at a one-month lag, leaving little room for a model to add forecasting value beyond a naive “assume no change” rule. The growth cycle, by contrast, offers a more genuine prediction problem while still being grounded in a well-established driver of factor rotation.

A Random Forest classifier was trained on twelve macro-only features drawn from FactSet’s economic data integration, which spans roughly 2.5 million data series, including lagged PMI and its change, the Conference Board Leading Economic Index, the OECD Composite Leading Indicator, the BAA-10-Year credit spread, the 10-Year/3-Month yield curve, and measures of housing, payrolls, market returns, the Fed Funds rate, consumer confidence, and M2 growth. Critically, PMI is used only as a lagged input feature, not as the target definition applied in real time, which prevents look-ahead bias at the level of any single observation since the feature values used to predict a given month’s regime were genuinely knowable at that time.

A second, separate discipline guards against look-ahead at the level of the model itself, distinct from lagging any single input: The model is retrained every three months on an expanding window that only ever fits on data available at each point in time, so that the parameters generating any given prediction could actually have existed then, rather than having been estimated using years of data that had not yet occurred. Expanding rather than rolling the window also has the practical benefit of preserving the relatively scarce Contraction-regime months rather than letting them age out over time.

Chart 1: RF Feature Importances

Source: FactSet

The model was trained on data from July 1963 to December 1985, validated out-of-sample from January 1986 to December 2002, and deployed from January 2003 onward, well before any FactSet QFL data enters the picture. Out-of-sample, the Random Forest achieved 91.7% accuracy, modestly below the naive lag-1 persistence baseline of 94.6%.

However, when used to time a long Momentum position, the model-based approach delivered an Information Ratio of 0.37 versus 0.31 for the naive baseline, indicating that even a modest edge in classification can translate into a more meaningful edge in economic significance.

In deployment, from 2003 onward, the model’s hit rate has been 90.8% overall, 92.1% in predicted Expansion months and 87.7% in predicted Contraction months. As of the most recent reading, PMI stood at 54.3 and the model’s prediction for the regime is Expansion.

Feature importance is concentrated: Lagged PMI and the OECD Composite Leading Indicator together account for roughly 63% of the model’s predictive weight, averaging 44.6% and 18.5% respectively across deployment checkpoints, with the Leading Economic Index a distinct third at 12.9%, ahead of payrolls growth at 5.2%. The yield curve and market return contribute almost nothing, at 2.6% and 1.3%, respectively, a reminder that not every intuitive macro indicator earns its place in the model.

That concentration is not static. Re-fitting the same expanding-window Random Forest at each of the 95 retrain checkpoints between January 2003 and July 2026 shows PMI_lag1 and the OECD Composite Leading Indicator holding the top two positions throughout, but their combined share moves with the cycle: it widens heading into the 2007-2009 financial crisis and again through 2020 as the model leans hardest on its two strongest leading indicators exactly when the regime is under the most strain, and it narrows in calmer stretches as payroll growth and the Leading Economic Index pick up a larger share of the weight.

The yield curve and market return never climb out of the bottom two importances in any period, which rules out the possibility that Chart 1 is an artifact of a single unstable fit.

Chart 2: ISM Manufacturing PMI With Predicted Regime (1948-2026)

Source: FactSet

Chart 2 shows this pattern across the full history: the shaded bands trace the model's predicted regime against the raw PMI level from 1948 onward, making visible both the multi-decade stability of the Expansion/Contraction calls and the handful of episodes, such as the slow 2007-2009 grind, where the model's colour band lags the PMI crossing 50. 

Chart 3: Information Ratio of HML and Momentum by Growth Regime (Pre- and Post-2007)

Source: FactSet

Chart 3 splits Fama-French HML (Value) and Momentum returns by realized PMI growth regime, before and after 2007, the point at which each premium’s unconditional, regime-blind form began its now well-documented decay (“momentum is dead” and “value is dead” narratives). Read through the growth regime, the story is better described as a reversal than a disappearance: In the 2007-2026 era, Momentum’s Information Ratio moves from 0.48 in Expansion to -0.34 in Contraction (a spread of 0.82), and HML moves from 0.08 in Expansion to -0.38 in Contraction (a spread of 0.46). Momentum’s spread is markedly wider than its 1963-2007 counterpart (0.37); HML’s spread holds roughly steady (0.44 pre-2007 vs. 0.46 post-2007) even as its unconditional premium collapses. The regime-conditional premium has not shrunk alongside the unconditional one: For Momentum it has widened outright, and for HML it has held up even as the blended, regime-blind version disappeared. This is the empirical basis for expecting a regime-aware overlay to add value: The signals it tilts toward not only differ by regime, but do so in a period that postdates the well-known decay of both premia. 

Chart 4: Deployment Predictions Versus Realized PMI (2003-2026)

Source: FactSet

Chart 4 plots this deployment-era record directly, overlaying the model's predicted-regime bands on realized PMI from 2003 onward so the misses behind the hit-rate figures above can be read against the specific months in which they occurred.

The three-month retrain cadence used throughout this study is a design choice, not a constraint of the model itself, so it is worth checking that the results are not an artifact of that particular setting. Re-running the deployment predictor with monthly retraining on the same expanding window and the same twelve features produces a marginally higher hit rate: 91.9% versus 90.8%, but a higher regime flip rate as well at 5.3% versus 4.6%.

The two cadences disagree on the predicted regime in only 7 of 283 deployment months (2.5% of the sample), and every disagreement resolves within a month or two as the slower cadence catches up. The three-month default therefore appears to trade a small amount of accuracy for meaningfully more stable regime calls, rather than a one-sided improvement foregone for convenience.

Timing the Composites

The mechanism for incorporating the regime signal into OWE is intentionally conservative. At each monthly rebalance, the model’s forecast for the coming month is retrieved, and the historical signal return series used to estimate Max IR expected returns and covariances is filtered to include only those historical months where the realized regime matched the predicted regime.

The optimization itself is otherwise unchanged: The same Max IR objective, the same style factor and turnover constraints, and the same Factor Mimicking Portfolio-based risk and return estimation described in the original article.

This is a discrete, regime-conditional re-estimation of inputs rather than a continuous blend of regime states, which keeps the approach transparent and auditable, at the cost of a harder edge at each regime transition.

• Universe: US Large Cap (Financials and REITs were excluded from the universe, consistent with the original study)

• Structure: The same twelve QFL signals across Sentiment, Quality, and Value composites, unchanged from the original study. Group-level bounds and the turnover cap were set wider than the original configuration for this initial run, a deliberate choice to first confirm the regime-aware mechanism behaves sensibly before tightening constraints in a subsequent iteration.

• Period: The backtest window runs from April 2009 to July 2026, 208 monthly periods, with the Regime-Aware strategy carrying an additional warm-up of roughly 27 months to ensure a minimum number of Contraction-regime observations before its first rebalance

Three strategies were compared over this common window:

1. Equal Weight: The benchmark every other approach must outperform.

2. Max IR: The unconditioned middle ground, using the same Max IR objective and estimation approach as the original article, with no regime information.

3. Regime-Aware Max IR: The new candidate, in which the historical inputs to Max IR are filtered to periods matching the predicted regime before each rebalance.

Table 1: Backtest of Three Strategies, April 2009 to July 2026

Source: FactSet

The headline result is that timing the composites with the regime signal improves the Information Ratio, from 0.75 for Equal Weight and 0.85 for unconditioned Max IR to 0.93 for Regime-Aware Max IR, while also reducing volatility, maximum drawdown, and Expected Tail Loss relative to Equal Weight. This is consistent with the broad thesis that conditioning composite weights on the macro cycle adds value beyond static or purely risk-based weighting.

The result is not, however, uniform once the analysis is disaggregated. The longest drawdown for Regime-Aware Max IR, at 35 months, remains meaningfully worse than unconditioned Max IR’s 24 months, even though it is well short of Equal Weight’s 44 months. From a career risk perspective, a regime model that occasionally lingers on the wrong side of a slow-moving transition can extend an underperformance streak, even as it improves average risk-adjusted returns.

At the composite level, the picture is similarly mixed: Equal Weight actually produces the strongest Quality composite in isolation, with an Information Ratio of 0.60 versus 0.47 for unconditioned Max IR and 0.41 for Regime-Aware Max IR, suggesting the regime overlay does not help (and may modestly hurt) every building block of the portfolio, even where it helps the whole. 

Chart 5: Information Ratio by Regime

Source: FactSet

Chart 5 makes this split explicit. It shows each strategy's annualized Information Ratio computed separately within Expansion and Contraction months, so you can see whether Regime-Aware Max IR's advantage holds within each regime or is concentrated in one side of the cycle. In fact, it leads neither split outright: Equal Weight has the higher Expansion-month Information Ratio, and unconditioned Max IR has the higher Contraction-month Information Ratio, with Regime-Aware Max IR a close second in both.

Its full-period advantage in Table 1 is therefore not explained by dominating either regime in isolation. It is more consistent with the strategy performing best around the regime transitions themselves, as the case studies later in this article illustrate. 

Chart 6: 24-Month Rolling Regime Prediction Accuracy

Source: FactSet

Chart 6 tracks this same accuracy on a rolling 24-month basis, showing how much the hit rate implied by the confusion matrix above has varied through time rather than holding at a constant 91.5%.

Table 2: Backtest of Strategies and Composites, April 2009 to July 2026

Source: FactSet

Table 3: Backtest of Signals, April 2009 to July 2026

Source: FactSet

A single hit-rate figure can hide an asymmetry that matters more than the average. This figure differs slightly from the 90.8% deployment hit rate quoted earlier because it covers a different, narrower window of the same prediction series: 272 months from December 2003 to July 2026 (the span actually used by the OWE backtest) versus the full 283-month deployment history from January 2003 to July 2026 (used in the regime signal section above).

The underlying model and predictions are identical in both cases; only the date range differs. Breaking this record into a confusion matrix shows 93.8% overall accuracy across 272 months, with recall of 94.9% in actual Expansion months against 90.8% in actual Contraction months, and precision of 96.4% when the model calls Expansion against 87.3% when it calls Contraction.

In both directions, the model is more reliable when it predicts Expansion than when it predicts Contraction. It misses actual contractions more often than it misses actual expansions, and a fair share of its Contraction calls turn out to be wrong. Given that Contraction is also the regime in which the composite-level Quality result was weakest, this is the kind of detail that is easy to lose in an aggregate accuracy number and is worth keeping in view. 

Chart 7: Risk Contribution by Strategy

Source: FactSet

Source: FactSet

Source: FactSet

Chart 7 disaggregates that risk contribution by portfolio building block. The overlay does not uniformly raise or lower any one group’s share of risk. Value’s risk contribution under Regime-Aware Max IR swings the most sharply of the three strategies, including brief spikes above 100% around 2016-17 and again in 2020, while its Quality and Sentiment risk contributions move within a narrower band that more closely tracks unconditioned Max IR. This is consistent with a mechanism that reallocates risk toward whichever composite the regime prediction favors at a point in time, rather than one that permanently overweights a single building block. 

Chart 8: 24-Month Rolling Information Ratio by Strategy

Source: FactSet

Chart 8 shows the Information Ratio advantage is not constant through time. All three strategies move together through the major drawdowns in the backtest, most visibly the sharp decline into 2019-2020, since they draw on the same underlying signals.

The separation between Regime-Aware Max IR and the other two strategies widens and narrows across different multi-year stretches rather than holding at a fixed margin. Investors should expect the overlay’s benefit to be episodic rather than a steady, period-by-period uplift. 

Chart 9: Average Signal Weight by Predicted Regime

Source: FactSet

Source: FactSet

Chart 10: Actual Weight by Strategy

Source: FactSet

Source: FactSet

Source: FactSet

Chart 10 complements this by showing each strategy's actual realized weight to a composite over the full backtest, rather than the average tilt by predicted regime shown in Chart 9.

The clearest illustration of what the regime overlay is trying to do comes from looking directly at the three most recent actual regime shifts in the backtest window, rather than at averages across the full period. Rebasing each strategy to 100 twelve months either side of the transition date shows Regime-Aware Max IR tracking unconditioned Max IR closely for most of each window, and it then diverges around the shift itself in the direction the mechanism is designed to produce.

This is the most visually persuasive evidence of the overlay earning its keep, but by construction it is also a small and cherry-picked sample of three events out of many months in the backtest and should be read alongside the full-period tables above rather than in place of them. 

Reading The Results with Care

The purpose of Chart 9 is to make the timing mechanism visible rather than to take it on faith. It compares the average signal weight the optimizer assigns in months predicted to be Expansion against months predicted to be Contraction. If the mechanism is working as intended, weight should tilt toward the composites and signals with the stronger historical regime-conditional Information Ratios shown in Chart 3, most notably the divergence in Momentum and Value performance between regimes.

Investors considering this type of overlay should look for that tilt to be economically sensible before trusting the aggregate Information Ratio improvement.

It is also worth being candid about what the historical factor evidence does and does not support. As Chart 3 shows, the regime split of HML and Momentum is a deployment-era pattern, not a full-sample one. The Value pattern fits a risk-based reading of HML as a rough proxy for distress risk since the highly levered, cyclically exposed businesses that dominate the value side are hit hardest as growth deteriorates. Quality shows little regime dependence at all over the full sample.

A regime-timing overlay therefore draws support from Momentum, Sentiment, and Value in the deployment era specifically (rather than from a dependable Quality-in-Contraction effect), and the composite-level results in this study are broadly consistent with that reading.

Summary

The result that matters most for whether this kind of overlay is worth building is not the backtest itself but what sits behind it. Both the Momentum and HML premia, whose unconditional forms have decayed toward zero since 2003, resurface as cleaner, wider swings once split by growth regime, with the regime-conditional spread larger in the 2003-2026 deployment era than in 1963-2003. A regime overlay is only worth the added complexity if the premium it is timing has not simply vanished. The deployment-era evidence here says it has not; it has reorganized around the growth cycle.

Layering a disciplined macro regime prediction onto the same Optimal Weights Engine framework used to weight signals within composites and then using it to condition (rather than simply average) the historical inputs to Max IR has produced a higher Information Ratio, lower volatility, and a smaller Expected Tail Loss than either Equal Weight or unconditioned Max IR over the April 2009 to July 2026 backtest. That improvement, however, is not free of trade-offs: the longest drawdown lengthened relative to unconditioned Max IR, and at least one composite, Quality, performed better without the regime overlay than with it.

As with the original study on signal weighting, the conclusion is not that one approach dominates in every respect, but that the differences are worth understanding, monitoring, and stress-testing before they are relied upon in production.

Two extensions follow directly: a systematic search across regime predictor model families and feature sets, since the predictor was deliberately kept simple and is the likeliest lever for a materially better outcome, and a systematic search across the group-bound and turnover-cap space, set loose here to first validate the mechanism.

Set against both, the quality of the underlying signals still matters more than how they are timed or bounded, and no amount of predictor or constraint refinement substitutes for improving the signals themselves.

I am thankful for feedback from Todor Bilarev and Ivan Vratzov in preparing this article.

This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.

¹ Svallin, J. “A Practical Approach to Weighting Signals.” FactSet, 2026.

² Hodges, P., Hogan, K., Peterson, J. R., & Ang, A. (2017). “Factor Timing with Cross-Sectional and Time-Series Predictors.” The Journal of Portfolio Management, 44(1), 30-43.

³ Hoffstein, C. “Style Surfing the Business Cycle” and “Macro and Momentum Factor Rotation.” Newfound Research, 2019.

⁴ MSCI. “Adaptive Multi-Factor Allocation.” Research Insight, October 2018.

⁵ Sheth, A., & Lim, T. (2017). “Fama-French Factors and Business Cycles.” Working paper, Saint Mary’s College of California. Available at SSRN: https://ssrn.com/abstract=3082577.