Improving the Six Indicators with the Tripwire Evidence

Analysis date: 2026-10-10. Method: mapped all 42 tripwires against the six indicators (RPI, LIG, FSM, RPS, DRG, CSG), then ran the three tests the mapping suggested. Tests: ~/workspace/valuation-model/work/analysis5.py (reproducible).

What the 42 wires decompose into

• 18 are already ingredients of the six (or near-duplicates): curve/Sahm/IP/permits/payrolls (RPI), claims/temp help (LIG), Baa-Aaa/NFCI (FSM), CAPE/Buffett/allocation/concentration/margin (DRG, RPS, CSG).
• 9 are slow confirms: quarterly delinquencies, debt service ratio, bankruptcy filings. Coincident-to-lagging; they confirm stress after it exists. They belong on the board, not inside probability models.
• 8 are market microstructure: 200dma, breadth, VIX level and term structure, 5-day drops. Timing conditioners, not cycle measures.
• 7 are genuinely new information the six do not contain: HY spread momentum (the divergence wire), BDC/private-credit market stress, job-openings momentum, real-pay squeeze, Subchapter V bankruptcies, SLOOS level, KBRA defaults.

The four blind spots the comparison exposed

1. The model layer measures credit in slow monthly grades; the live crack is in daily market prices. FSM reads 0/100 today while HY spreads have widened +55bp off their low in six weeks and the BDC basket is -19.7% from its high. The board and the model currently contradict each other.
2. No private-credit channel exists in any of the six. It lives only on the board (Fitch 6.30%, KBRA 3.5%, BIZD). For the bubble Michael ranks first, the model layer is blind.
3. A household cash-flow squeeze is invisible to RPI and LIG. Real pay is -0.2% YoY and small-business Sub V filings are +46% YTD, but unemployment hasn't moved, so labor-based gauges read calm. This "quiet deterioration" regime only reaches the models when it reaches payrolls.
4. Fragility and triggers remain disconnected. DRG (95) cannot price Michael's named trigger class (a single-name/IPO event). No formula fixes that; the board's divergence and BDC wires are the closest early sensors.

Test results

Test A — add credit features to RPI (1959-2026). Adding Baa-Aaa level and 3m widening: in-sample AUC 0.875 → 0.897, but out-of-sample AUC (train ≤1999, test 2000+) 0.827 → 0.822, i.e. no real improvement. The widening coefficient even flips sign (conditional on level, widening marks recessions already underway). Verdict: do not change RPI's formula. Its actual weakness is interpretive: after the Dec 2024 re-steepening, its probability is structurally low (8.5%). Fix the display, not the model: show the post-inversion clock (22 months since re-steepening vs. the 1-11 month historical lead window) next to the probability.

Test B — FSM's widening leg. Adding more legs the naive way (FSM2 with HY + SLOOS legs averaged in) diluted the gauge (2008-09: 80 vs 100) and still read 0 today, because monthly-mean HY erases the late-August low. The correct fix is measurement, not ingredients: make the widening leg max(Baa-Aaa monthly widening, HY daily widening from its 6-month low). That variant (FSM3) reads 33 today instead of 0, keeps 100 in Sep 2008, 50 in Dec 2015 and Jun 2022. Caveat: the HY leg is only computable from Oct 2023 locally (ICE licensing limits FRED's HY history to ~3 years), so the improvement is justified by the live 2026 case and construction logic, not a long backtest. Verdict: adopt.

Test C — job openings for LIG. Rejected by the data. The openings ratio collapsed during recessions, not before (Dec 2007: 0.59 already; 12m change into the peak only -0.09; into Feb 2020: +0.06). Its level is also not portable across eras (1.05 was normal in 2006; 1.01 today follows a 2.04 peak). Verdict: keep as a tripwire only; LIG unchanged.

Recommendations (priority order)

1. FSM widening leg → max(Baa-Aaa, HY daily) (reads 33 today). Apply the same swap to CIS's credit component in the VFS/CIS layer, which has the identical monthly blindness (its credit component is 0 today while HY is +55bp off the lows).
2. CSG: add the earnings-gap line (top-10 cap share minus earnings share: +12pp in 2000, +9-11pp end-2025, ~+3pp Oct 2026) so the gauge distinguishes supported from unsupported concentration instead of reporting the level alone.
3. RPS: publish the uncertainty band. Expected real return -0.7%/yr with a ±1 SD band of -5.4% to +3.9% (residual SD 4.6pp). Even the optimistic edge is below the ~6.5% long-run average; the band makes the precision honest.
4. RPI: display fix only (post-inversion clock beside the probability). No formula change; the credit-feature test failed out-of-sample.
5. No change to LIG or DRG. LIG's candidate addition failed its test. DRG stays a pure valuation/positioning gauge; breadth and household stress are timing and cycle information, and blending them in would repeat the mistake the two-layer design exists to avoid.
6. Architecture stands: six slow indicators + the 42-wire board as the fast sensor layer. The one structural addition worth making later, if the private-credit wires keep deteriorating, is promoting BIZD/BKLN stress from the board into FSM as a fourth leg — flagged now, adopted only when there is enough history (BIZD starts 2013) to grade it honestly.

Research note by Michael Davis (Taslos). Research, not investment advice.