The last article ended in a no. But “direction is not predictable” does not mean “nothing is measurable”. The right question is narrower: in a price series, what has memory?
One measurement, two opposite answers
Autocorrelation answers exactly that: how much today tells you about tomorrow, the day after, ten days on. Compute it twice on the same data — once on returns (signed, so: direction) and once on absolute returns (unsigned, so: size).
^GSPC, ten years, lags 1 to 20 sessions.
The grey line — direction — hovers around zero and goes nowhere: 0.048 at lag 5, −0.048 at lag 10. That is noise.
The teal line — size — starts at 0.37 and is still 0.20 twenty sessions later. That is structure, and it decays slowly.
The same split holds elsewhere: BTC-USD 0.159 against −0.049; AAPL 0.232 against −0.059. Not a quirk of one market.
What it means
Violent days come in clusters. After a large move, the next session is more likely to be large too — while which way it goes is anybody’s guess. Markets forget direction almost immediately and remember magnitude for weeks.
So if only one thing can be measured, measure the one with a memory.
Why this suits you better than it suits a fund
A fund monetises magnitude by selling options, running volatility books, holding hundreds of names so the law of large numbers can do its work. None of that is open to you.
But you can use the same fact far more simply: it tells you where an order can still be reached, how far a stop must sit to survive ordinary noise, and which target price actually gets touched often enough to name. Three answers, enough for a plan, and not one of them needs a direction.
What follows from it
The question changes from “where will price go” to “how far does it travel, and how often”. The second one has an answer, and the answer is enough to build a plan: where to rest an order so it can actually be reached, how far a stop must sit to survive ordinary noise, what target price is genuinely touched often enough to be worth naming.
None of those needs a direction. All of them need a distance.
Which raises the next question: if the past teaches us something, why not tune the parameters on the past as well?