I have a composite risk score I look at most mornings. Some blend of liquidity, credit spreads, cross-asset volatility, positioning, the usual ingredients. For a long time it just sat there as a number between roughly zero and one hundred, and I would nod at it and then go do whatever I was going to do anyway. The score was informative and completely useless at the same time, because it never told me to change my exposure. It was a thermometer with no thermostat attached.
The fix is boring and it is the whole game. You have to decide, in advance, what each region of the score means for how much risk you carry. That decision is the threshold band. Get it right and the score starts doing work. Get it wrong and you will either whipsaw yourself to death at the boundary or set the levels so wide that the signal never fires.
Let the distribution set the bands, not your gut
The first instinct is to pick round numbers. Risk-off below 30, risk-on above 70, neutral in the middle. It feels clean and it is almost always wrong, because your score does not live on a nice uniform scale. A composite built from spreads and volatility spends most of its life bunched in a narrow range and only occasionally sprints to the extremes. If you anchor bands to round numbers, you might discover the score is above 70 maybe twice a decade, and you have effectively built a risk-on signal that never turns on.
So set the bands from the score's own history. Take a long window, ideally spanning at least one real stress event, and look at the percentiles. A reasonable starting point is something like this:
- Risk-off when the score sits in roughly its worst 20 to 25 percent of readings historically.
- Risk-on when it sits in roughly its best 20 to 25 percent.
- Neutral for everything in the middle, which is most of the time, and that is fine.
The exact percentiles are a taste and a risk-tolerance question, not a law. What matters is that you are cutting the distribution where the observations actually are, so each band gets used often enough to matter and rarely enough to mean something. A quick sanity check I run: count how many days a year each band would have triggered over the backtest. If risk-off shows up on two days total, the band is too extreme. If it shows up half the year, it is too loose to be a warning.
One caveat. If you rebuild the percentiles on a rolling window, the bands move under your feet, and a score that was risk-off yesterday can be neutral today without the underlying reading changing at all. That is usually not what you want. I prefer fixed bands recalibrated deliberately, maybe once a year, rather than bands that quietly drift. Slow and known beats fast and invisible.
Add hysteresis so you stop flip-flopping
Here is the failure mode that ruins most band systems. The score is hovering right at your risk-off line. It dips below, you cut exposure. Next reading it ticks back above, you add it back. Then below again. You are now trading the noise around a single threshold, paying spread and fees every time, and your portfolio is doing an impression of a nervous person at a light switch.
Hysteresis fixes this, and it is the same trick a thermostat uses so your heating does not click on and off every thirty seconds. You use two lines instead of one. The level that pushes you into risk-off is stricter than the level that lets you back out. Say you enter risk-off when the score drops below its 20th percentile, but you do not leave risk-off until it climbs back above its 30th. That gap in the middle is a no-mans-land where you simply hold your current state and do nothing.
The practical rules are:
- Define an entry threshold and a separate, more forgiving exit threshold for each band.
- When you are inside a band, ignore the entry line entirely. Only the exit line can move you out.
- Size the gap between entry and exit to the score's typical day-to-day wiggle. If the score routinely moves a few points on noise, your gap needs to be wider than that wiggle or the hysteresis buys you nothing.
The cost of hysteresis is that you are always a little late. You enter risk-off a bit after the deterioration starts and you leave it a bit after conditions recover. You are paying that lateness on purpose to buy far fewer round trips. In my experience the trade is worth it, because the whipsaw losses from a hair-trigger boundary quietly add up to more than the lag ever costs you.
Size to the band, do not trade the trigger
The last mistake is treating the band crossing like a light switch. Risk-off, so go to cash. Risk-on, so go all in. Binary positioning off a composite score is fragile, because the score is an estimate with error bars, and you are betting the whole book on it being exactly right at the moment it crossed a line.
Map each band to a target exposure instead, and let the band set the size. Something like full or near-full risk budget in risk-on, a reduced budget in neutral, and a defensive floor in risk-off. You are not going to zero even in the worst band unless your mandate demands it, because a composite score is a probabilistic read and it will be wrong sometimes. Leaving a residual position in risk-off is what keeps a false alarm from costing you the recovery.
You can go further and scale continuously within a band, so exposure glides as the score moves rather than jumping at the edges. That smooths turnover even more, though it is more to maintain and you should confirm the extra machinery actually improves the result before you commit to it. Either way, the score is now steering position size, which is the only thing it was ever good for.
A concrete end-to-end workflow I would trust: build the composite, pull several years of history including at least one drawdown, set entry and exit percentiles per band with a hysteresis gap wider than the daily noise, map each band to a target exposure with a non-zero floor, then backtest the turnover and the drawdown you would actually have lived through. When I wire scores to positioning inside Blockcircle, that turnover number is the one I stare at longest, because a signal that is right on direction and horrible on trading frequency will still bleed you out.
None of this makes the score smarter. It just stops the score from being a decoration. The bands, the hysteresis, and the sizing are the plumbing between a number you look at and an exposure you actually carry, and that plumbing is where most of the real value hides.