Regime Filters Need Units: A Standardized Index Is Not a Thirty-Day Expectation

A regime filter is only as well specified as the unit of the series it reads. That sounds like a truism until two series arrive on the same data portal carrying the same Units label, both drift slowly, and both end up wired into a threshold that was calibrated on only one of them. The Chicago Fed National Financial Conditions Index and the Cboe Volatility Index are the pair that most often gets handled this way, because both are treated on trading desks as shorthand for how tense the environment is.

The conclusion first, before any of the readings. A threshold calibrated on an option-implied volatility index cannot be carried onto a standardized index, and a threshold calibrated on a weekly release cannot be evaluated on a daily bar, because the weekly series has not produced a new number since Friday. Neither claim rests on a forecast or on a view about direction. Both follow from what the two publishers say the numbers are.

What follows works through the published definitions row by row, then through five consecutive weekly readings as released, then through the size of the move those readings contain. The readings are small, and that is the point rather than a weakness in the example: a filter that treats any directional drift as a regime change will fire on a move of this size.

Same Units Label, Different Quantity

Five fields separate the two series. They are listed here in the same order as the rows in the diagram below, and each one is taken from the publisher rather than from convention.

First, what the number is. The Federal Reserve Bank of Chicago states in its description of the NFCI that "The NFCI is a weighted average of a large number of variables (105 measures of financial activity) each expressed relative to their sample averages and scaled by their sample standard deviations." The Cboe methodology, by contrast, states that "The VIX Index measures the 30-day expected volatility of the S&P 500 Index," and the note field on the FRED page for VIXCLS describes it as market expectation of near term volatility conveyed by stock index option prices. One is an aggregation of many indicators against their own histories. The other is a calculation on a single option surface.

Second, the stated basis of the scale. The Chicago Fed is explicit: "The NFCI and ANFCI are each constructed to have an average value of zero and a standard deviation of one over a sample period extending back to 1971." That single sentence supplies the entire interpretive frame for the series. It tells a reader what one unit means, where the neutral point sits, and how to convert a raw move into a statement about how unusual it is. The volatility index scale is established by a different route, through the calculation set out in the Cboe VIX methodology document, which describes the index as "a financial benchmark designed to be an up-to-the-minute market estimate of the expected volatility of the S&P 500® Index." These are two different ways of fixing a scale, and neither one licenses a threshold borrowed from the other.

Third, the release cadence. The FRED series page for NFCI prints the frequency as Weekly, Ending Friday. The VIXCLS page prints Daily, Close. A backtest that joins these two on a daily index and then evaluates a joint condition each session is, for four sessions out of five, re-reading a number that has not moved. Whether that matters depends on what the filter does with it, which is a question worth settling before the backtest rather than after.

Fourth, the horizon the number describes. The conditions index is released as a weekly update, and the publisher pages do not state a horizon it looks forward over. The volatility index, by the methodology sentence quoted above, describes the following thirty days. A filter that compares today's volatility index reading against today's realised range is setting a thirty-day quantity against a one-day quantity. There are defensible reasons to do that, but the comparison is not an identity and should not be presented as one.

Fifth, the Units field itself. Both series pages print the identical string: Index, Not Seasonally Adjusted. That field is doing no work here. It distinguishes an index from a rate or a dollar amount, and it does not distinguish a standardized deviation from an option-implied expectation. A data pipeline that keys unit handling off that field will treat the two as interchangeable, and nothing in the field will object.

Reading Five Weekly Releases as Published

The chart below shows the conditions index and its adjusted counterpart for five consecutive weeks, plotted against a zero axis with no truncation, and with the one standard deviation line drawn in so that the size of the move can be judged against the scale the publisher defined.

Taken as released, the headline index read -0.546 for the week ending 2026-08-07, then -0.551, -0.555, -0.560, and -0.564 for the week ending 2026-09-04. The adjusted index over the same five weeks read -0.587, -0.585, -0.584, -0.585, and -0.588.

Two features of that sequence are worth separating. The headline index moved in one direction in every one of the four steps, for a total change of 0.018. The adjusted index did not: its total range across the window is 0.004, and its least negative reading falls in the middle of the window rather than at either end. On the Chicago Fed reading of sign, where "Positive values of the NFCI have been historically associated with tighter-than-average financial conditions, while negative values have been historically associated with looser-than-average financial conditions," both series sat on the loose side throughout, and the headline series drifted slightly further onto that side while the adjusted series did not.

The two are not required to move together. The Chicago Fed states that the adjusted index "removes the variation in the individual indicators attributable to economic activity and inflation before computing the index," so a window in which one drifts and the other does not is a property of the construction rather than an anomaly to be explained. What this window cannot establish is which of the two better describes the environment, or why the drift occurred. Five observations of a weekly series do not support that kind of statement, and a piece that made it would be reading far more out of the ANFCI release than the release contains.

What a Move of 0.018 Can and Cannot Support

Because the index is constructed to have a standard deviation of one, the arithmetic here is unusually direct. A total change of 0.018 is 1.8 percent of one standard deviation. That is the whole of the headline move across five weekly releases, expressed in the only unit the series actually has.

This is where a filter specification either holds or quietly fails. A rule that flips state on any directional move would have flipped on this window, four steps in a row, and would have reported a trend. A rule that requires a displacement of half a standard deviation before changing state would have recorded nothing at all, correctly, because nothing of that size occurred. The two rules disagree entirely, and the disagreement has nothing to do with the market and everything to do with the threshold being expressed in a unit that matches the series or not.

Direction is also cheaper than it looks. With four steps, a strictly monotone sequence is not a rare event on its own. The temptation is to read four consecutive steps as confirmation, but in a slow weekly index built from overlapping inputs, persistence of that kind is closer to the default than to evidence, and treating it as a signal is the same error as reading a single funding print as stress, which is the pattern set out in the note on quarter-end funding spikes and the calendar.

One further comparison makes the mismatch concrete. The volatility index closed at 17.20 on 2026-09-15 according to the FRED VIXCLS page, a different series on a different cadence covering a different horizon. That number is not plotted above and is not commensurate with the conditions readings. Any rule that combines the two is a joint rule, and a joint rule needs its own specification.

Concrete Framework

The steps below run in order, and the order matters: each one closes off a way the next one can go wrong. They apply to any published series a desk wants to use as a state variable, not only to the two discussed here.

1. Name the unit. Write down, in one line, what one unit of the series is, before any threshold is chosen. For a standardized index the answer is one sample standard deviation. For an option-implied index the answer comes from the calculation in the methodology. If the line cannot be written, the threshold that follows is arbitrary.
2. Name the horizon. State the span the number describes, taken from the publisher, not the span the position is intended to cover. When the two differ, record the difference rather than papering over it. A thirty-day expectation and a one-week aggregate are both legitimate inputs and neither is a daily reading.
3. Set the threshold in the series own units. A standardized index is thresholded in standard deviations. A threshold carried over from an option-implied volatility index has no meaning against it. This is the step where most borrowed rules break, and it is also the cheapest step to check.
4. Require persistence. Demand a move that survives several consecutive releases rather than one print, and state the number of releases in advance. For a weekly series, a persistence requirement stated in trading days is a category error and should be restated in releases.
5. Write the invalidation line first. Record, before the filter goes live, the reading or the behaviour that would retire it. A filter without a pre-committed retirement condition tends to acquire one after the fact, tuned to whatever has just happened. The same discipline applied to position limits appears in the discussion of how an initial requirement and a maintenance level interact.

What Would Invalidate This

Several things would undercut the argument above, and they are worth stating plainly rather than leaving implicit.

• A change in construction. The unit argument rests entirely on the standardization sentence the Chicago Fed publishes. If the index were rebased, or the sample period changed, the meaning of a 0.018 move would change with it, and every threshold expressed in standard deviations would need recalculating. The claim is a claim about the current published construction, not a permanent one.
• Evidence from a long window. If a properly constructed test over many years showed that moves of roughly 0.018 in this index reliably preceded some measurable change, the argument that such a move is too small to threshold on would weaken considerably. Five weekly observations cannot settle that question in either direction, and nothing above should be read as having settled it.
• A filter that does not use levels. The unit objection applies to level thresholds. A filter that uses the series as one input to a rank, or that compares each series only against its own history, is not exposed to the mismatch described here. For those designs the relevant question is different, and the five rows in the first diagram are less decisive.
• Revisions. The five readings above are the values as they appeared on the series pages at the time of writing. If a later release restates them, the specific arithmetic changes even though the framework does not, in the same way that a position list built from a stale filing misstates the present, as set out in the note on reading a quarterly holdings snapshot.
• A desk that never mixes the two. If a filter reads only one series and the threshold was calibrated on that series, none of the cross-series argument applies. The failure described here requires the borrowing, and plenty of specifications do not borrow.

The Case for Thresholding Anyway

There is a real counterargument, and it deserves better than a dismissal. Round-number thresholds on widely watched indices are coordination points. A level that many participants watch can matter because it is watched, independently of whether it is a unit-valid comparison, and a rule that keys off such a level may perform acceptably for that reason alone. Someone running such a rule is not making an arithmetic mistake so much as making a different kind of bet.

The reply is conditional rather than flat. If the rule is justified on coordination grounds, then the justification should be written down as such, and the rule should be tested against the behaviour it claims to exploit rather than against a story about financial conditions. What does not survive is the middle position, where a threshold is chosen because a number looks high, described as though it measured tightness, and then defended as a coordination effect once it stops working. Each of those three defences is testable, and they are not testable in the same way.

A second objection is more practical. Desks rarely have the luxury of a fully specified filter, and a rough state variable that is directionally right most of the time may beat no state variable at all. That is often true, and it remains true only while someone knows which of the five specification steps were skipped, because an unspecified filter cannot be debugged when it misfires.

Carrying This Into a Live Filter

The practical residue is small and checkable. Before a series becomes a state variable, its unit, its horizon, its cadence and its neutral point should each be written on one line, taken from the publisher rather than from habit. A threshold should then be expressed in that unit. Where two series are combined, the combination should be stated as a rule in its own right, with its own persistence requirement and its own retirement condition, rather than inherited from whichever of the two was calibrated first.

On the five weeks above, that discipline produces an unexciting answer: a drift of 0.018 in a series with a standard deviation of one, alongside an adjusted series that did not drift at all, does not support a change of state under any threshold worth running. An unexciting answer that can be checked against the published release is worth more than a confident one that cannot, and the check takes a few minutes with the publisher's own pages open.

This is a note on reading published series, not investment advice, and nothing above is a recommendation to buy or sell any security.