A composite macro score answers one question well and a second question not at all. It tells you how much deceleration the data is showing. It does not tell you what is causing the deceleration, and the two plausible causes call for opposite portfolio responses.
If activity is softening because demand has weakened on its own, the appropriate response is defensive and the appropriate duration of that response is long. If activity is softening because policy is more restrictive than the economy can carry, the response is still defensive but the exit is different, because the constraint is removable and the removal is at the discretion of a committee that watches the same data you do. A score that cannot separate those two is asking the investment committee to guess at the most consequential fork in the analysis.
The gap in a seven-model row
Look at what the Macro Risk Scorecard actually carries. The model tiles are labelled Blockcircle Labs at 41, GDP two quarter rule at 30, yield curve at 30, Sahm rule at 20, credit stress at 10, leading indicators at 39, and a combined score at 29, with a counter showing 0 out of 7 models at or above the elevated threshold of 60.
Read those labels for what they measure. Output, employment, credit pricing, the term structure, and a leading composite. Every one is a measure of outcomes. None of them is a measure of the policy stance that helped produce the outcomes, and none is constructed to isolate over-tightening as a distinct source of risk. That is not a criticism of the design, it is an observation about what the seven slots cover, and it is why the component described here is something you build alongside the module rather than something you switch on inside it.
The practical consequence is that a falling composite is currently uninformative about cause. You can read a regime label of SLOWDOWN and a macro health grade of C at 59 off the header and still have no basis in the score for saying whether the source is policy or demand.
What a policy error component actually has to measure
Be precise about the target, because the loose version of this idea produces a restrictiveness gauge rather than an error gauge, and they are different.
Restrictiveness is how tight policy is. Policy error risk is the probability that the current stance is tighter than the economy requires, which requires both a measure of the stance and a measure of whether the economy is absorbing it. A high policy rate in an economy running hot is restrictive and not an error. The same rate in an economy where transmission has already bitten is a different object.
So the component needs at least two limbs: how far the stance sits from a neutral estimate, and how much of that stance has already arrived in the real economy. A single-limb version will flag every tightening cycle as an error, which is the same as flagging none of them.
Four inputs, and how to get them onto one scale
The inputs below are all constructible from published series. The normalisation matters more than the choice of inputs.
- Stance distance. The real policy rate, deflated by an inflation measure you have fixed in writing, against a neutral estimate. Carry it as a band across at least three published neutral estimates rather than a point, because the spread across estimates is frequently wider than the distance you are trying to measure.
- Transmission progress. How much the stance has already reached borrowers. Lending standards surveys, effective rates on new credit, and the pace of credit growth. This limb is what stops the component from firing on the first hike of a cycle.
- Pace. The cumulative change in the policy rate over a trailing window, ranked against its own history. Speed of tightening is a distinct risk from level of tightening, because adjustment costs fall on borrowers who cannot refinance fast enough.
- The market's own vote. The shape of the expected policy path. When the curve of expected rates slopes down over the coming year, the market is saying the current stance is not sustainable. That is a genuine input and it is also the one most likely to double count, for reasons covered below.
Normalise each limb to a percentile of its own trailing history, using only information available at each historical date, then map to a 0 to 100 scale so the component is directly comparable with the existing model tiles. Percentile ranking rather than z-scoring, because these series are not symmetric and a z-score will make the tails behave strangely at exactly the readings you care about.
The double counting problem, which is the whole difficulty
Adding an eighth component to a seven-model row is trivial. Adding one that carries independent information is not, and the yield curve tile is where the trouble lives.
A term structure signal already encodes a market judgement about the policy path. If you build a policy error component with a large weight on the shape of the expected rate path and then place it beside a yield curve model, the two will move together, the composite will appear to have gained confirmation, and what has actually happened is that one piece of information has been counted twice. The counter of models at or above the elevated threshold is particularly vulnerable to this, because it is a count, and a count treats two correlated tiles as two independent votes.
The test is mechanical. Compute the correlation between your candidate component and each existing tile over the longest common history, using the same real-time discipline. If the correlation with the yield curve model is high, either drop the market path limb or reweight until the residual information is what you are actually adding. Report the correlation matrix in the specification document. It is the first thing a reviewer should be able to check and it is almost never included.
Apply the same test to the credit stress tile against your transmission limb, for the same reason. At the current reading of 10 and MINIMAL the two would look unrelated, which tells you nothing, because correlation estimated in a quiet period is estimated on the observations that matter least.
Using it for attribution, not as a trigger
The temptation once the component exists is to trade it. Resist that, at least initially, and use it for the job it was designed for.
The output belongs in the monthly pack as an attribution line beside the composite. When the composite deteriorates, the pack should state how much of the deterioration is accompanied by an elevated policy error reading and how much is not. That single sentence changes the shape of the committee discussion, because a policy-driven softening implies a different expected duration for a defensive posture than a demand-driven one, and it implies watching a different set of releases for the turn.
It also changes what you write when a position goes wrong. A defensive stance taken on a demand deterioration that turns out to have been a policy overshoot, subsequently reversed, is a specific and learnable failure. Without the attribution line, that outcome is recorded as the macro call simply being wrong, and nothing is learned.
The falsification test to agree in advance
Before the component enters the pack, write down what would prove it is not working, and get that signed off while nobody has a position riding on it.
My preferred test has two parts. First, does the component add explanatory power over the existing tiles for forward changes in the composite? Run the regression, report the incremental contribution honestly, and be prepared for it to be small. Second, and more important, does the attribution split produce different subsequent paths? Sort historical episodes of deteriorating composite readings into high and low policy error buckets and compare what happened to the composite over the following months in each bucket. If the two buckets look the same, the component is decoration regardless of how sensible its construction is.
The sample will be thin. Tightening cycles are infrequent and each one contributes a handful of genuinely independent observations, so the honest conclusion will often be that the evidence is weak. That is an acceptable place to land, provided it is stated in the document and the component is weighted accordingly rather than promoted to a trigger on the strength of a plausible story.