Reweight the machine, and it survives out of sample.

Everything the first study measured at pillar level is re-measured at component level, on longer data, against a harder test.

The component score histories come from the pipeline's own index engine — the same fixed-band and point-in-time percentile scoring that builds the published index, captured mid-run, never re-implemented. Three data additions extend the reach: the ICE BofA high-yield OAS series restored in full (weekly from December 1996, after the FRED licence cap left the machine's own OAS components with eighteen usable months); CFTC Commitments of Traders positioning back to 1986; and maximum-history market series — the S&P 500 to 1970, the All Ordinaries to 1984, the dollar index to 1971, the 10-year Treasury to 1962. Everything before December 2004 is genuinely out of sample for every claim the first study made.

The validation protocol is walk-forward: at each month the indicator's sign is re-learned from completed observations only (minimum sixty), the signal is that sign times the reading's percentile in its trailing ten years, and the out-of-sample record accumulates from the first qualifying month — roughly 2010 onward, up to 206 months. Each indicator's confidence score (0–100) is mechanical: 25% full-sample strength, 30% walk-forward retention, 20% sub-period consistency, 15% cross-era evidence, 10% sample depth. Without pre-2004 confirmation the ceiling is ~92 — no indicator gets to claim a century it hasn't lived through.

The withdrawn first filing described the bearish acceleration cell backwards. The machine's process is to say so, precisely, and move on.

The first study reported the risk-off state as a fall that is accelerating. The code's arithmetic says otherwise: the cell that loses money is a fall that is *decelerating* — the composite's 3-month change negative but shallower than a quarter earlier. That is an early-stage rollover: the machine has started easing off from a top, and the grind lower is still ahead (−2.84% per quarter, 29 episodes). An *accelerating* fall is the opposite case — capitulation mid-crash — and it leans contrarian-bullish (+1.72% per quarter over 94 episodes, 64% hit), consistent with the deep-fall and zero-crossing results. The corrected definitions are what this study validates and what any dashboard flag will carry.

Three findings carry over intact, now with harder evidence behind them: the composite is a rates instrument, the pillars cancel for equities, and the tails are asymmetric.

The equal-weight composite leads AU 10-year yield changes at 6–12 months (IC +0.25 to +0.27, both halves, survives an implementation lag) and is silent on equity direction (|t| < 1.1 everywhere) because its monetary pillar is procyclical while its economy pillar is contrarian — the average buries both. Regime shape survives: the AU tail warning (12-month collapse below −1.25 points) preceded −10.8% average six-month returns over 17 overlapping months; zero-crossings remain contrarian in both countries; deep US 3-month falls preceded +8.2% six-month returns against a +4.9% unconditional mean. The sector map also carries: info tech leads in weak-machine regimes on both exchanges, staples fade in recoveries, AU energy trades with the machine.

Column guide — Finding: the study-I claim. Evidence: its headline statistic (IC = Spearman rank correlation between the reading and the forward outcome, −1 to +1). Status: how study II re-graded it after component-level, out-of-sample and pre-2004 testing.

3,164 screen cells, kept only where the sign agrees in both halves of the sample. The strongest cells are not equity cells — they are real rates against gold, easing cycles against copper, labour against the Australian dollar.

The single strongest relationship in the entire study is the US real-policy-rate score against gold twelve months out: IC −0.59 (NW t −5.8), the same sign in both halves, and a +0.52 walk-forward IC. The direction matters and is not the naive one: *restrictive* real rates today have preceded gold strength over the following year — gold prices the easing that restriction eventually forces — while deeply negative real rates have marked the point where gold's run was largely behind it. Its silver and dollar-index analogues follow the same anticipation logic. The AU cash-rate 12-month-change score — the easing-cycle component — leads copper (+0.42) and WTI (+0.40) a year ahead: an Australian easing cycle is, in effect, a leading read on the global commodity cycle. US labour deterioration leads copper and the Australian dollar the same way. On equities the standouts are the US inflation-target score's 12-month change (+0.37 vs the S&P, stable halves) and the AU drawdown score against high-yield spreads (+0.43) — a market that has *not* drawn down is the credit-widening warning, which is the crowding story told through prices.

Column guide — Signal: the transform of the component score (level, or its 3/12-month change) and the asset outcome it is tested against, h months forward. IC: Spearman rank correlation between reading and outcome (−1 to +1; beyond ±0.30 is strong for monthly data). NW t: t-statistic with Newey-West correction for overlapping forward windows (|t| ≥ 2.0 ≈ 95% confidence, ≥ 2.6 ≈ 99%). IC 1st/2nd half: the same correlation measured separately in each half of the available window — matching signs mean the relationship is not one episode. n: months observed. Plain read: what the relationship says, in words, with its statistical confidence.

Signed weight grids over every component with 120+ months of history, scored on rank IC, selected for stability across a median split — never for the in-sample maximum.

Equal weighting is the wrong answer inside nearly every pillar, for the same reason it was wrong across pillars: components disagree in sign. The AU markets pillar equal-weighted reads *negative* against forward equity returns (−0.20); re-signed — drawdown and breadth inverted, volatility score positive — it becomes the strongest pillar composite in the study (IC +0.51 in each half, though on a 2013-onward window that deserves suspicion in proportion to its size). The US economy pillar's stable weighting collapses onto a single component: the inverse CPI-target score. The US monetary pillar wants the real-rate score inverted, the curve slope positive, and the long high-yield OAS score — the component the restored full-history series rebuilt — at full weight. The argmax column is reported beside every stable selection, and the sensitivity matrix perturbs every weight by ±0.25: the stable optima sit on plateaus, the argmax corners do not.

Column guide — EW IC: the pillar's predictive correlation with forward 6-month equity returns using equal component weights (the published construction). Reweighted IC: the same test using the stable signed weights shown. IC 1st/2nd half: the reweighted composite's correlation either side of the window's midpoint — the stability test that drove selection. n: months in the common evaluation window. Stable weights: signed component weights, normalised so the absolute values sum to 1; a negative weight means the component enters inverted.

Component-reweighted pillars, recombined on a signed grid, selected for cross-half stability, then sent through the walk-forward. This is the study's objective, delivered with its uncertainty attached.

The US optimised index — markets −0.25, economy −0.33, monetary +0.08, sentiment +0.33, over the reweighted pillar composites — carries an IC of 0.357 against forward six-month S&P returns, and the split-halves read 0.357 and 0.356: as close to invariant as this data can produce. Walk-forward, with the sign re-learned monthly and no knowledge of the future, it holds +0.41 over 175 out-of-sample months at a 62% hit rate. The AU optimised index concentrates in reweighted sentiment (+0.8) against the economy pillar (−0.2): IC 0.329 full, 0.342/0.334 by half, +0.30 walk-forward. One honest asterisk on both: the stable-selection step sees both halves, so the *selection* is in-sample even though the *weights* are validated out-of-sample by the walk-forward and the pre-registered sign structure from study I. The next upgrade is adoption: publish the optimised series monthly beside the equal-weight index, and let the live record become the test.

Column guide — Weights: signed pillar weights over the component-reweighted pillar composites (absolute values sum to 1; negative = inverted). IC full / 1st / 2nd half: rank correlation with forward 6-month equity returns, whole window and either side of its midpoint. Walk-fwd IC: the correlation achieved out of sample when the relationship is re-learned each month from past data only. WF hit: share of out-of-sample months on the right side of the median outcome. n: months.

Every indicator the two studies surfaced, walk-forward validated, with a mechanical confidence score from 0 to 100 — and the failures left in the table.

Ten indicators sit at the ceiling of 92: the real-rates complex against gold, silver and the dollar; the easing-cycle and labour reads on copper; the machine-to-yields cluster; the inflation-normalisation equity signal; the drawdown-to-credit warning; and both optimised indices. The rates cluster follows at 85–89. At the bottom, scored and kept deliberately: the AU monetary pillar's equity signal (confidence 41 — its walk-forward IC is negative despite a +0.34 in-era IC) and the US economy contrarian (46 — near-zero out of sample). A register that only lists winners is marketing; this one is the evidence base.

Column guide — h: forward horizon in months. IC: full-sample Spearman rank correlation between the indicator and the outcome. Walk-fwd IC: the correlation achieved out of sample, sign re-learned monthly from past data only (positive means the indicator works live, whatever the raw sign). Hit: share of out-of-sample months on the right side of the median outcome (50% = coin flip). OOS months: length of the out-of-sample record. Confidence: mechanical 0-100 score — 25% strength, 30% out-of-sample retention, 20% sub-period consistency, 15% cross-era evidence, 10% sample depth; ceiling ≈ 92 without pre-2004 confirmation. Tier: Core ≥ 85, Usable 60-84, Rejected < 60. Plain read: the relationship in words with its live record.

Five discrete states with episode-level track records — too rare for walk-forward, too important to drop.

The AU tail warning remains the sharpest alarm the machine issues: seventeen overlapping months of 12-month collapse beyond −1.25 points, followed by −10.8% average six-month returns against a +2.7% unconditional mean. The corrected early-rollover flag is the working risk-off trigger. The two contrarian events — zero-crossings and accelerating falls — sit on the other side of the book: by the time the machine confirms a downturn, positioning data says the crowd is already out.

Column guide — Episodes: months on which the trigger was active (overlapping where forward windows overlap). Mean/Median fwd: average and median forward return of the target after the trigger. Hit: share of episodes that moved the expected way. Uncond. mean: the same asset's average forward return over ALL months — the baseline any event must beat. Window: first and last trigger dates. Plain read: what the alarm means and how literally to take it.

Component histories reach back decades further than the published index. Rebuilt into extended composites, they let the modern era's claims face data that was never part of forming them.

The verdicts are uncomfortable and therefore valuable. The economy-pillar contrarian signal — one of the first study's cleanest modern-era results — reads +0.03 against the S&P across 1949–2004: it does not exist before the QE era, and its register score reflects that. The AU monetary pillar's equity relationship flips sign outright (−0.15 pre, +0.37 post): in a high-rates world, easing meant recession; in a low-rates world, easing meant multiple expansion. The relationships that *do* hold across eras are the quiet structural ones — the economy composite against 10-year yields, and the credit-spread reads. The lesson for the index's design: equity-directional weightings must be treated as era-conditional and re-learned, which is precisely what the walk-forward already does; rates and credit weightings can be trusted with longer leashes.

Column guide — IC pre-2004: rank correlation measured on data that predates the published index entirely (a true out-of-sample test, since every claim was formed on 2004-26 data). IC 2004-26: the modern-era correlation. n pre/post: months in each era. Verdict and Plain read: whether the relationship survives outside the era it was discovered in, and what that means for how far to trust it.

CFTC net speculative positioning, 1986 to date, answers the first study's open question: why are the machine's confirmations contrarian?

Because the crowd gets there first. The machine's 3-month change leads next-month speculative flows in S&P futures (ρ +0.25) — by the time the composite crosses zero, positioning has already adjusted, which is why the crossing marks exhaustion rather than onset. The conditioning table makes it explicit: a weak machine with crowded speculative shorts is the best S&P cell in the sample (+5.7% per quarter, 88% hit over 34 months); the same weak machine with longs still crowded offers nothing comparable. Crowded oil longs fade even in strong machines. Positioning does not replace the machine — it explains which of the machine's states are tradeable and which are already spent.

Column guide — Cell: the conditioning state (speculative-positioning tercile from CFTC net non-commercial futures positions as % of open interest, crossed with the machine's level). Result: mean forward 3-month move of the asset in that state, with the share of months positive. n: months in the cell. Plain read: what the state implies and with what reliability.

The index carries two instruments with two jobs: the equal-weight composite as the state variable, and the optimised overlay as the predictive series beside it.

The design conclusions. The published equal-weight composite is unchanged — it remains the honest description of conditions, the regime map, the rates dial and the carrier of both standing alarms (the tail warning and the corrected early-rollover flag). The optimised series runs beside it with weights frozen as tabled — component weights inside each pillar, signed pillar weights across them — re-examined only on a stated schedule, never tuned to the month. The long OAS history is bound into the pipeline so the monetary pillar's credit components carry real memory, and the register — including its rejected rows — is the standing evidence base, re-scored as each month's outcome completes.

The limits, stated plainly. The stable-selection step is still selection — the walk-forward validates the chosen weights but the choosing saw the whole sample; only live months cure that. Several component windows are one cycle deep (AU markets reweighting rests on 2013-onward data; AU monetary's curve and credit-spread components begin in 2015). Forward windows overlap and Newey-West corrects imperfectly at twelve months. The gold and silver results lean on a metals era that may not repeat. COT is weekly, futures-only, and non-commercial is an imperfect proxy for the marginal speculator. And every confidence score above 46 in the equity column belongs to the modern monetary era — the pre-2004 test is the standing reminder that the machine's equity edge is conditional on the regime it was learned in.