Portfolio Allocation by Signal

Allocation as Signal

Most market commentary treats portfolio allocation as opinion. This post treats it as a signal-processing problem: given the current state of every quantitative input mplot tracks, what allocation does the data imply? The output is a deterministic function of the signals — not a forecast, not a recommendation, just the position the model holds when forced to decide.

The exercise is useful for two reasons. First, it forces the analytical apparatus to commit to a number. Second, it makes the model's behavior observable: when readers return next month, they can see what changed and why, rather than reading another opinion piece.

The methodology and triggers are explicit below. Replicate, criticize, or replace as needed.

Methodology

The allocation rule is intentionally minimal:

1. Universe: Ten assets tracked by the multi-factor scanner — broad equity (SPY, QQQ, IWM, EFA, EEM), bonds (TLT, HYG), gold (GLD), and crypto (BTC, ETH).
2. Inclusion: Only assets with positive multi-factor composite score are eligible. The composite weights momentum (40%), carry (30%), and value (30%).
3. Sizing: Allocations are proportional to composite signal strength, then capped per asset class to control concentration: 30% bonds, 25% high-yield credit, 25% equity, 8% per crypto position.
4. Cash reserve: 15% base, plus 5% if recession probability exceeds 25%, plus 5% if the liquidity z-score is below -0.5. Both adjustments are currently active, producing a 25% reserve floor.
5. Overflow: When position caps bind, residual capital is iteratively redistributed across non-capped positive-signal assets. Anything that cannot be placed accrues to cash.

The rule does not consider correlation, factor crowding, transaction costs, or tax. Those are second-order. The first-order question is which assets the cross-sectional signal currently favors, and the methodology answers that with one composite per asset.

What the Model Holds

As of April 30, 2026, four assets clear the positive-composite threshold: TLT (+0.62), BTC (+0.71), HYG (+0.51), and ETH (+1.07). All other assets — including SPY, QQQ, GLD, and EFA — register zero or negative composite scores and receive no allocation under the rule.

The resulting portfolio is heavily defensive. Long-duration treasuries take the maximum 30% allocation, supported by strong carry (+1.04) and historically cheap valuation (+1.51). High-yield credit fills 25%, driven entirely by the highest carry score in the universe (+2.16) — HYG's 5.9% yield offers more cushion against capital risk than any other asset class currently tracked. The remaining risk-taking is concentrated in 8% positions in BTC and ETH, where the multi-factor model registers strong momentum (+1.63 and +1.92) and positive value relative to five-year price history. Cash absorbs the residual 29%.

The absence of equity exposure in this allocation is the single most counterintuitive feature. SPY rallied roughly 8% over the past six weeks, which is reflected in its positive momentum score (+0.28). But that single positive factor is overwhelmed by SPY's negative value score (-0.81), reflecting an estimated CAPE near 41 — well above any reasonable historical anchor. The composite (-0.11) is therefore slightly negative, and the rule excludes it. This is a feature, not a bug: a methodology that buys equities at every CAPE level provides no information; one that occasionally refuses to is at least falsifiable.

Factor Decomposition

The multi-factor breakdown clarifies why the model holds what it holds. ETH and BTC score positive across two of three factors (momentum strongly, value mildly) and negative on carry — the standard crypto signature. TLT and HYG are mirror images: weak momentum, strong carry, and (for TLT) strong value. Gold, conventionally framed as a hedge in the current environment, scores negative on all three factors: momentum has rolled over after the early-2026 highs, carry is structurally negative, and value remains stretched against its five-year mean. The model assigns it zero weight despite the macro narrative.

The decomposition also highlights the disagreement between short- and long-window momentum. The standalone momentum scanner (63-day) ranks SPY first; the multi-factor model penalizes SPY's valuation and ranks crypto first. Neither view is correct in isolation. The composite is the resolution.

Triggers

The allocation is not static. The trigger dashboard lists the six signals whose state changes would force a re-allocation, along with the threshold and direction.

Two triggers are particularly close to firing. First, the Sahm Rule sits at 0.20, well below its 0.50 threshold — but the recession-probability composite (28.3%) is already past the 25% threshold that activated the recession-related cash adjustment. A 20-25 basis-point increase in the unemployment trend would push the rule meaningfully closer. Second, the liquidity z-score at -0.85 is deep in tightening territory. A reversal toward zero would not change the allocation; only a sustained move above +0.5 (an Easing regime confirmation) would prompt a substantial increase in equity and crypto exposure at the expense of cash.

The remaining triggers are farther out. A re-inversion of the 10Y-3M curve from +0.67 would call for extending bond duration. A move in crypto Fear & Greed from 29 to above 75 would prompt taking profits in BTC and ETH. A reversal in gold's multi-factor score from -1.07 to above +0.5 would re-introduce a 10% gold position. These are not predictions — they are the conditions under which the model would change its mind.

Limitations

This methodology is one model, not the model. Five points of honest caveat:

The composite weights (40/30/30) are heuristic, not optimized. A different weighting could materially change rankings, particularly for assets where momentum and value disagree.

The vol-based caps (8% crypto, 25% equity, 30% bonds) are imposed by hand, not derived. They reflect risk-management intuition but are not the output of any optimization.

The CAPE-driven negative value score for equities assumes mean reversion to a historical anchor. If valuations are structurally higher in the AI/productivity-growth regime, the model is overweighting that signal.

The cash reserve logic uses two binary thresholds rather than a continuous function. A continuous version would be more elegant but offers little practical difference at current levels.

Most importantly: this allocation has not been backtested. The rule is observable going forward; whether it produces returns above a 60/40 benchmark is an empirical question that requires several years of out-of-sample data to answer.

The point of publishing it is not to claim performance. It is to commit to a methodology in advance, so that future readings of mplot can be evaluated against a real prior.