# If the framework missed 2002, how much can we trust it for 2026?

*1 Oct 2026. Draws on `backtest-2002.md`, the live vote record in `dataroom/register/regime_votes.csv` (Jan 2003 to Sep 2026), and the 1 Oct 2026 journal. The framework and the journal are unchanged; the suggestions in section 5 are for the next journal entry and a possible framework revision.*

## 1. Short answer

**Trust it to describe where we are and to organise the thinking. Do not trust its headline vote to call a turn, which is the very thing 2026 asks.**

The framework has four layers, and they did not all fail in 2002:

| Layer | What it does | 2002 result | Confidence for 2026 |
|---|---|---|---|
| 1. Decompose the yield | Splits the 10-year into expected inflation, real rate and term premium | Worked: it showed the fall came from real yields and term premium | **High.** This is accounting, not forecasting |
| 2. Structural drivers | Scores seven slow forces at −1, 0 or +1 | **Right:** −5 of a possible −7, pointing to lower yields | **Moderate.** It passed its one test, but one test is little evidence |
| 3. Regime vote | 3 of 5 headline metrics in a regime's bands, 3 months running | **Wrong:** A1 B2 C2 in May 2002, no A call from Nov 2001 to Dec 2006 | **Low at turning points**, reasonable mid-regime |
| Ranges (journal method) | Applies past 12-month changes to today's level | **Wrong:** the outcome (3.37%) was below every historical case | **Low in the tails.** Treat ranges as too narrow |

So the 2002 failure is mostly a failure of the **regime vote** and of **history-based ranges**. Both are built to describe a regime once it has arrived. Neither was designed to detect one arriving.

## 2. The evidence is worse than one missed year

The live data room runs the same vote on every month-end from Jan 2003. Its record:

| Period | What happened to yields | What the vote said | Verdict |
|---|---|---|---|
| 2002–2008 | Disinflation regime begins; 10-year falls from 5% to below 4% by Nov 2008 | No A call until **Oct 2015**, 13 years after the turn | Missed the turn |
| 2015–2020 | Deep in Regime A | A for 35 months (Oct 2015 to Nov 2020, with gaps) | Right, but late |
| Aug 2021 to May 2022 | Inflation regime begins: core PCE 3.9% rising to 5.6%; 10-year from 1.30% to 2.85% | **A for 10 months running** | Wrong at the turn, the other way |
| 2024–2026 | Higher real rates: 10-year 4–5.3% | B for most months (Apr–Aug 2024, Feb 2025 to Jan 2026, May–Aug 2026) | Right, mid-regime |
| 2003–2026 | Includes 2021–23, when core PCE inflation peaked at 5.6% | **Never called C in 285 months** | Cannot see Regime C |

Two turning points, two misses, in opposite directions. The pattern is consistent: **the vote reads the regime that just ended.** In 2021 it called disinflation because real yields and the term premium were still negative, levels the 2026 bands score as A. The framework's own Layer 3 table describes "real 10-year falling or negative while nominal yields rise" as the Regime C signature, but the vote turns that into an A vote whenever the real yield is low.

## 3. Why the vote fails at turns

1. **Fixed level bands carry the calibration of their era.** The 2026 bands were set from the 2010s and 2020s. In 2002 a 2% term premium and a 3% real yield were normal, so the vote read a disinflationary decade as B or C. In 2021 negative real yields were the legacy of QE, so it read the start of an inflation regime as A.
2. **Market levels move last at a regime change.** Levels confirm a regime after prices have repriced; the direction of change and the structural drivers move first.
3. **The vote has gaps exactly where turns happen.** It casts no policy vote when the Fed is hiking, so it was half-blind through 2022. It casts no TIPS vote when the real yield is above every band, as in 2002 (3.16%).
4. **Three months running is strict.** In 2002–03 the direction-aware variant reached three A votes four times but never three months in a row.

## 4. What this means for 2026 right now

The same weaknesses are visible in today's readings (month-end Sep 2026):

- **The vote has stopped calling.** Sep 2026 is **A1 B2 C0, no call**, after B in May–Aug. Two of the five inputs cast no vote:
  - the 10-year TIPS yield is 2.91%, above every band (the 2002 problem);
  - fed funds rose 25bp over three months, and a hike earns no policy vote (the 2022 problem).
- **The fiscal input points the wrong way.** Interest/receipts is 21.1%, above the B band, yet it votes **A**, because it is falling slightly. The C rule needs it both above 20% and rising.
- **The other inputs sit near band edges.** The 5y5y (2.36%) and the term premium (0.85%) are both in the B band. But in May 2002 the 5y5y also read B, and its band was set in the 2010s.
- **The structural drivers point up.** The framework's 2026 starting score is about **+3 to +5** (savings vs investment +1, fiscal +1, globalisation +1, demographics and safe-asset demand 0 or +1). In 2002 Layer 2 was the part that got the direction right. Today it points to higher yields, toward B with C risk, not A.
- **Inflation is not at target.** Core PCE is 3.0% (Aug 2026), above the B description of "about 2–2.75%".

Read together, today looks more like a regime in transition, or a B regime with C risk building, than a settled B. That is the situation where the back-test says the vote is least informative. It is also the situation the framework exists for.

## 5. How to use the framework, given this

1. **Treat the vote as confirmation, not a forecast.** A call from the vote tells you a regime has been in place for months. Lack of a call does not mean nothing is happening.
2. **Lead with Layer 2 and with direction.** At each review, record the driver score, how it changed since last time, and the 12-month direction of the 5y5y, the real yield and the term premium. These moved first in 2002.
3. **Cap regime probabilities near 55%.** One framework, one partial back-test and two missed turns do not justify more. The journal's data-based B 55% is at that cap; the gut B 50% is inside it.
4. **Widen the tails of the yield ranges.** In 2002 the outcome fell outside the 5th–95th range. The journal's data-based 10-year range of 4.25–6.50% at 75% leaves 25% outside, which is right in spirit, but write down explicitly what happens to the portfolio at 3.5% and at 7%.
5. **Size positions for being wrong.** The decision rules in framework section 7 (separate the view from the position; use carry as cushion) matter more than the probabilities.
6. **Score the calls.** The first `forecasts.csv` rows resolve on 1 Oct 2027. Scoring them against outcomes is the only evidence that will raise confidence in the framework itself.

## 6. Changes that would raise confidence (not yet made)

| Change | Why | Effort |
|---|---|---|
| Back-test 2021–22 in the same format | The most recent turn, and the one most like 2026 (inflation, fiscal expansion, Fed behind the curve) | A few hours (copy `backtest_2002.py`) |
| Back-test 1994, 1999–2000 and 2013 | Bond rout, late-cycle hikes, taper tantrum: overlays versus regime changes | A few hours each; TIPS data starts 1999 |
| Add a hiking rule to the policy vote | Today a hike casts no vote; "Fed hiking and 30-year rising faster" could count as C | Small code change in `derive.regime_votes`, plus a framework revision |
| Score the real yield relative to its own 10-year history (as in the direction-aware variant) | Removes the era calibration of the bands | Variant already exists in the back-test code |
| Treat a negative or falling real yield with rising nominals as C | Matches the framework's own C description; would have flagged 2021 | Small code change |
| Show the Layer 2 score next to the vote in the report | Keeps the leading layer in view | Small report change |

Each of these is a framework revision. Make them in a new framework version, test them on 2002 **and** 2021–22, and only then use them live. A rule tuned to one episode will misfire on the next.

## 7. Bottom line

The framework is a **discipline**: it says what to measure, how to split a yield move into its causes, and which readings would change the view. That part is sound, and the data behind it is checked and reproducible. Its **mechanical regime call** is a lagging indicator, wrong at both turns it has faced (2002 and 2021). For 2026 the right level of confidence is **moderate in the description, low in any single regime call**: probabilities near 50%, wide tails, and positions that survive the regime you did not expect.