Zyberno Technical Note · Methodology & Backtest

A methodology & validation note documenting a live Zyberno tool. This is not part of the Brina Gap framework research; it is supporting technical work.

The Zyberno Market Valuation Score: Construction, Backtest, and the Limits of Optimization

By Fabio Brina · Zyberno · June 2026 · Data window 1986–2026

Abstract. We document the construction and historical behaviour of the Zyberno Market Valuation Score, a single 0–100 reading that blends four widely-followed macro gauges — the Buffett Indicator and Shiller PE (valuation), and the yield curve and credit spread (recession risk). Rebuilt monthly from 1986 to 2026 across four U.S. recessions, the score is strongly inversely related to subsequent 10-year S&P 500 returns (r = −0.78) but is not a reliable recession-timing tool — it was low before both the 1990 and 2008 downturns. We further show that optimizing the component weights overfits: the in-sample-optimal and least-squares-optimal weightings both fail out-of-sample, while a simple, robustness-validated weighting (20/40/30/10) generalises. Adding a market-sentiment signal adds no predictive power. We therefore present the score as a transparent long-run valuation gauge — useful for setting return expectations, not for market timing — and publish it live with full methodology.

1. Motivation

Individual valuation gauges are familiar but noisy, and each tells only part of the story. The Buffett Indicator and Shiller PE say how expensive equities are; the yield curve and credit spread say whether the financial system is flashing stress. Our aim is modest and explicit: combine them into one transparent number, then ask honestly what that number has — and has not — been able to do. In an environment where generic market commentary is increasingly automated, a gauge earns trust through reproducible construction and candid validation, not through a confident headline.

2. Components and Data

The score uses four monthly series. The common window begins in 1986, set by the shortest input (the Baa credit spread). Quarterly inputs are carried forward as-of (no look-ahead). Recession dates are the NBER business-cycle reference dates (FRED series USREC); forward returns use the S&P 500 monthly price.

GaugeDefinitionSourceHistory
Buffett IndicatorTotal market cap ÷ GDPFRED NCBEILQ027S ÷ GDP1951–
Shiller PE (CAPE)10-yr cyclically-adjusted P/EShiller / multpl.com1871–
Yield curve10-yr minus 2-yr TreasuryFRED T10Y2Y1976–
Credit spreadMoody's Baa minus 10-yr TreasuryFRED BAA10Y1986–

3. Construction

Each gauge is mapped to a 0–100 sub-score by where it sits in its historical range (0 = cheap or calm, 100 = extreme): the Buffett Indicator over 75–200%, the Shiller PE over 10–40, the yield curve over +1.5% to −1.0%, and the credit spread over 1.5–5.0%. These thresholds are set by economic judgment and held fixed — they are not fitted to outcomes. The composite is a weighted average, kept to a 60% valuation / 40% recession-risk split. The weights are the one element we examine empirically in §5.3.

4. Backtest methodology

We compute the score at each month-end from 1986 to 2026 (485 observations) and study three things: (i) its level in the run-up to each NBER recession; (ii) its correlation with subsequent S&P 500 price returns at 1-, 3-, 5- and 10-year horizons; and (iii) whether re-weighting the components improves prediction out-of-sample (train 1986–2005, test 2006–2026). Return windows overlap, so significance is overstated and we read magnitudes as indicative; returns are price-only.

5. Results

5.1 It is not a recession timer

The score gave a clear warning before the 2001 dot-com bust — but was low before both the 1990 recession and the 2008 global financial crisis, when valuations were only middling and credit was still calm.

Recession (NBER peak)Score 12m beforeScore at onsetPeak (prior 24m)Warned?
Jul 1990311731No
Mar 2001784680Yes
Dec 2007492950No
Feb 2020574859Partial (pandemic)

The recession signal lives almost entirely in one component: the yield curve, which inverted roughly 19, 34 and 25 months before the 1990, 2001 and 2008 recessions (and not before the 2020 pandemic shock). We therefore show the yield curve as its own gauge and do not market the composite as a recession predictor.

5.2 It predicts long-run returns

What the score is good at is exactly what a valuation gauge should do — set expectations for the next decade.

Horizoncorr(score, forward annualised return)
1 year−0.07
3 years−0.37
5 years−0.52
10 years−0.78

By component, the 10-year signal is carried by valuation — Shiller (−0.88) and Buffett (−0.54) — with a modest contribution from the yield curve (−0.33) and essentially none from the credit spread (+0.09), which is a coincident stress gauge rather than a return predictor. Splitting history at the median score, the top half of readings preceded about +5.5%/yr over the next decade, versus +10.8%/yr from the bottom half — roughly half the return from an expensive starting point.

5.3 Optimizing the weights overfits

It is tempting to fit the weights for maximum predictive power. We did — exhaustively — and the result is a cautionary tale. The in-sample-optimal weighting collapses onto ~100% Shiller (r = −0.88 in-sample) but, fit on 1986–2005 and tested on 2006–2026, its out-of-sample correlation falls to +0.07. An unconstrained least-squares fit — the best possible linear combination over all weights and signs — scores a flattering +0.91 in-sample and then −0.28 out-of-sample, i.e. actively wrong on unseen data. The naïve balanced weighting generalised better (−0.41) than either "optimised" version.

The robust choice is therefore not the optimum but the weighting that holds up across eras. We adopt Buffett 20% · Shiller 40% · yield curve 30% · credit 10%: it keeps the 60/40 valuation/recession identity, applies the only durable lessons (favour Shiller over the redundant Buffett; favour the yield curve over the return-irrelevant credit spread), matches the baseline's full-sample correlation (−0.78), and is far more stable across the two halves of history (−0.81 / −0.67 vs −0.87 / −0.41).

5.4 Sentiment adds nothing

We tested whether a market-sentiment signal (a price-momentum and volatility proxy for a Fear & Greed-style index, since real Fear & Greed history spans only ~2 years) improves the score. Its correlation with forward returns ranged from −0.06 to −0.13, and the optimizer assigned it zero weight. Sentiment is a days-to-weeks timing signal, not a decade-ahead valuation signal, so it stays out of the composite and is shown separately on the live dashboard.

6. The score through history

025507510019901995200020052010201520202025
The Zyberno Market Valuation Score, monthly, 1986–2026 (weights 20/40/30/10). U.S. recessions shaded; the tint marks the elevated zone (60+). Today's reading of 72.6 sits in the 97th percentile of the last 40 years (range 9.5–79.9).

7. Limitations

We state these plainly. (1) Forward-return windows overlap, so the correlations are autocorrelated and their statistical significance is overstated; we treat magnitudes as indicative. (2) The window contains only four recessions — low power for any recession claim. (3) Returns are price-only; total returns would be ~2%/yr higher across the board, leaving the relationship intact. (4) Sub-score thresholds use the full historical range, a mild look-ahead in levels (not in the weight tests, which are split out-of-sample). (5) The sentiment input is a proxy, not the live Fear & Greed index. (6) The analysis is U.S.-only and uses a single train/test split. We deliberately did not fit nonlinear, interaction or regime-switching terms: the linear out-of-sample failures above show such flexibility would overfit, not generalise.

8. Conclusion

The Zyberno Market Valuation Score is best understood as a transparent, long-run valuation gauge. Its honest job is expectation-setting: when it is high, the next decade has historically paid roughly half what it paid from low readings. It is not a market-timing or recession-prediction device, and we do not present it as one. At the time of writing the score sits near the top of its 40-year range, which the historical relationship associates with below-average forward returns — a sober reading, offered with its uncertainty attached. Future work will ask whether this macro regime conditions the stock-level signal of the Brina Gap framework.

References

See the live Zyberno Market Valuation Score →