Which currency questions are actually answerable?

01 · The shape of the problem

The market hides which way it will move. It telegraphs how rough the ride will be.

Before building anything, ask the price history one question: does what happened over the last few days tell you anything about today? Ask it twice. Once about the direction of moves, once about their size.

The blue bars sit at zero on every pair. Yesterday’s direction tells you nothing about today’s. One bar on the rupee panel does dip below zero, since the rupee tends to give back a little the day after a move, but it is too small and too short-lived to build on. The models confirmed that later. The orange bars run clearly positive everywhere: wild days follow wild days, calm days follow calm days.

That single difference decides all three verdicts below. Q1 has no answer in this data. Q2 and Q3 do.

02 · Q3 · Planning rate

Plan next month around today’s rate, not last month’s average

Businesses need one exchange rate agreed in advance: a budgeting rate, a pricing rate, or a rate written into a contract. What that number is really guessing at is next month’s average rate. Almost nobody tests it as its own question.

We compared three ways of setting it at every month-end across 104 months, 2017 to 2026. Score it by asking how far off the guess was, on average, as a percentage of the average that actually happened. Smaller is better.

Average miss, %. 104 months per pair, lower is better

A · last month’s average B · today’s rate C · small statistical model

Guess C, the statistical model, ties guess B on every pair. The win comes from a better starting point. The modelling adds nothing. The whole test fits in a few lines: at each month-end, record what each method would have guessed, then compare it with the average that actually happened.

Source: scripts/14_period_average.py, condensed

03 · Q2 · Roughness

Size is forecastable where direction is not

“Volatility” just means how much the price jumps around. A calm week versus a wild one. Knowing next week will be wild says nothing about which way prices go, but it does tell a business how much safety margin to leave.

Two forecasters raced each other one week ahead, over 426 separate weeks per pair. The do-nothing guess assumes next week is as wild as the recent past. The volatility model is a textbook recipe called HAR. It blends yesterday’s choppiness with last week’s and last month’s, weighting each by what has worked historically.

Source: scripts/10_turbulence_har.py, condensed

The model cuts forecast error roughly in half on every pair, 51-57% less, measured by the scoring rule statisticians use for volatility forecasts. That rule punishes under-warning about a storm harder than over-warning about one.

Next-week forecast error. 426 weeks per pair, lower is better

Error score is the volatility scoring rule described above. Units matter only for comparison. Monthly results in A.3.

Look at the third row of every block. Stacking machine learning on top of the simple recipe made forecasts worse every time at this horizon. The extra complexity cost accuracy and bought nothing.

This tells you how much cushion to leave, and nothing about which way the rate moves. In a week forecast to be wild you buy protection sooner and price in a bigger buffer. Put less at stake. In a week forecast calm, run the reverse.

04 · Q1 · Direction

Nothing beat “the next period repeats the last one”

Fifteen approaches were graded on direction: classic trading rules, statistical models, machine-learning models, and combinations of all three. Every one raced the laziest guess available. Assume the next period does what the last one did.

Accuracy is scored out of 100, so “+6.4” means the best model was right about six more times per 100 guesses than the lazy one. Every advantage here came in smaller than the uncertainty around it. Any of them could be luck.

Extra correct guesses per 100 vs lazy guessing. 322 predictions per pair, final unseen test period

A model can look fine on accuracy and still be dangerously overconfident

Some models state a confidence with every guess (“70% sure it goes up”). We checked whether those confidences are honest: when a model says 70%, is it right about 70% of the time?

Logistic regression sits near the line, so its stated confidence means something. Gradient boosting is far off it. When it claimed “90% sure” it was right barely half the time. If you act on a stated probability, this chart matters more to you than any accuracy table.

The Fed’s words measure cleanly and add nothing

The second dead end got tested rather than assumed. All 99 FOMC statements were scored hawkish to dovish by counting phrases, a method any reader can check. The scorer passes its sanity check: the 2022-23 rate-hike era reads hawkish, the 2020 emergency cuts read dovish.

Fed into the models across the same 402 weeks, the scores changed nothing. The input was rejected. The experiment stays in the repo as a documented negative, because a measurement can be valid and still add no forecasting value, and both halves of that are worth publishing.

05 · The most transferable result

When a forecast looks too good for its field, audit the timestamps first

Early on, one model looked 70% accurate at predicting the next day. Implausibly good. The cause was a timestamp mismatch: Yahoo stamps its currency prices earlier in the day than it stamps the dollar index. Join the two on the same date and “today’s” dollar move quietly carries information about tomorrow’s euro move. The model was peeking.

The check that caught it uses correlation, a score from −1 to +1. Near 0 means no connection. Near −1 or +1 means strongly connected.

Source: notebooks/01_direction.ipynb, section 1

After the one-day shift, the 70% “accuracy” collapsed to ordinary guessing. Every number on this page comes from the fixed data. Of everything in this project, finding that bug is the part most likely to be useful to someone else.

06 · The close

What ten years of public data rules out, and what it supports

07 · How it was tested

Baselines named before models were built