Diren Kumaratilleke
IVSignal

NIV.

Thrust as the foundation for regenerative economies — the physical mechanics of capital formation, measured against cumulative friction.

1

What I Inverted

If this holds, central banks and allocators get an 18-month stress signal 41.71% orthogonal to the yield curve.

Recession indicators read bond-market sentiment. The 10Y/3M spread reads what bond traders expect the Fed to do, and expectations are reflexive to policy: quantitative easing moves the spread directly, and moves investment only through the real economy.

I measure the physical mechanics instead. Regime health is the velocity at which capital forms where margins compound — regenerative capital formation — against the cumulative friction impeding its circulation. Regenerationism is the philosophy; NIV is its first instrument.

NIV_t = ( u_t · P_t² ) / ( X_t + F_t )^η u = tanh( 1.0·ΔG + 1.0·ΔA − 0.7·Δr ) Thrust P = ( Investment × 1.15 ) / GDP Efficiency X = 1 − ( TCU / 100 ) Slack F = 0.4·s + 0.4·max(0, r−π) + 0.2·σ Drag

Thrust is the kinetic impulse mobilising capital, bounded by tanh. Efficiency is the regenerative-capital kernel, squared so that productive allocation is rewarded nonlinearly. Slack is capacity headroom. Drag is a weighted penalty over yield inversion, positive real rate and rate volatility. Thrust is discounted by friction rather than added to it.

2

What It Shows

The protocol first, because it is what a reviewer checks. Expanding-window walk-forward over 504 months, 1970–2024. The model retrains every 5 months. Warm-up runs through 1983. At every step only data up to t is used to predict the label at t + h; the target is never visible to the fit that predicts it. A 15-year rolling window was tested and rejected — it outputs zero probability 78.2% of the time.

Eight public FRED series and nothing else. GPDIC1 (real private domestic investment) drives thrust and efficiency. M2SL (M2 money stock) drives thrust. FEDFUNDS drives thrust and drag. GDPC1 (real GDP) drives efficiency. TCU (capacity utilisation) drives slack. T10Y3M and CPIAUCSL drive drag. USREC supplies validation labels only and is never an input. Every weight above is published, so a macro researcher with a free FRED key can rebuild the series from this page.

HorizonEnsemble AUCBrierOpt. F1
3 mo0.77020.09490.3471
6 mo0.74440.11600.2875
12 mo0.82430.09720.3590
18 mo0.85380.08910.4545

At 18 months NIV reaches 0.8538 against roughly 0.72 for the 10Y/3M spread on the same window and the same NBER labels. Accuracy improves with horizon rather than decaying — a contraction in regenerative capital formation takes about eighteen months to reach headline output.

The ensemble crosses its 35% alert threshold 7 times in 504 months, a 98.5% false-alarm filter; its individual layers cross 474, 230 and 50 times. All seven survivors are nameable events.

Correlation with the Fed spread is 0.7635, leaving 41.71% orthogonal variance. Gini importance puts the regenerative-capital term at 0.9328 and the spread at 0.0298 — the model picked the thesis unprompted.

3

What It Doesn't Show

The effective sample is about eight, not 504. The window holds 504 months but only around eight US recessions. Every interval here rests on that number, and no amount of monthly resolution changes it. This is the most important limitation on the page.

Out-of-sample but historical. The protocol prevents lookahead within the window, but the whole history was known to me when the framework was designed. There is no live forward track record.

Confidence bands widen to (0, 1) at the GFC peak, when the three ensemble members disagreed most. The model flags its own instability rather than averaging it away — correct, and an admission that at the moment of greatest interest it was least certain.

Methodological feedback was received from Ben Bernanke during development. That is feedback on method received while building — not an endorsement of the framework, the results, or the conclusions, not a review or an approval, and not to be cited as any of those.

4

Sources