Predicting the price is hopeless. Predicting the uncertainty isn't.

Open predictions

Forecasts already committed to git, waiting on the calendar — grouped by ticker so all three arms' ranges can be compared at a glance. The commit history is the proof they were made before the outcome was knowable — see it here.

Volatility the real positive result

Same discipline as the price arms above: a model forecast scored in dollars-equivalent units against a trivial baseline, on identical rows. Here the target isn't next month's price, it's the realized volatility over the next 21 trading days (annualized standard deviation of daily returns) — and the baseline is persistence of volatility: "the next 21 days will be as calm or wild as the last 21." The model is a simple EWMA estimator, not a neural net — the point is to show a real edge exists with an honest, well-understood method, not to show off a bigger one.

Skill by ticker

All 15 tickers are positive — a consistency check, not just a headline average that could be one lucky ticker away from evaporating.

Predicted vs. realized volatility over time

Sampled roughly weekly for readability; skill and RMSE above are computed from the full daily series, not this downsampled view.

Why volatility is predictable when direction isn't

Daily stock returns are close to a random walk — the sign of tomorrow's move carries almost no information about the sign of the next move, which is the whole reason persistence is such a tough baseline for direction. But the size of moves behaves completely differently: calm periods tend to stay calm, and turbulent periods tend to stay turbulent, a well-documented empirical regularity called volatility clustering — it's the entire reason GARCH-family models exist in econometrics. An exponentially-weighted moving average of recent squared returns exploits exactly this: recent turbulence is weighted more heavily as a forecast of near-future turbulence, and it works, consistently, across every ticker tested here.

Live track record committed before the outcome

Scored only once 21 trading days have actually passed. Coverage and interval score are always shown together — coverage alone is trivially gamed by a wide-enough range.

Calibration fixed, and shown honestly

An "80% interval" is a promise: the real outcome should land inside it about 80% of the time, no more and no less. This scoreboard shipped with a real defect — its intervals only covered 67-68% of the time, both arms, a promise it wasn't keeping. Coverage is never shown alone below, because it's trivially gamed: predict a wide-enough range and you can "cover" everything at the cost of the range being useless. Width and interval score (a score that penalizes both misses and needless width) are shown alongside every coverage figure, before and after.

The cause turned out not to be the initially-suspected volatility-regime shift (coverage by volatility tercile came back nearly flat) but overlapping-window residual autocorrelation shrinking the effective calibration sample size. The fix — split-conformal prediction with the nonconformity score normalized by each row's own trailing EWMA volatility — brought both arms inside the pre-registered success band, with width up only 1.15–1.22x and interval score improving for both arms, which rules out "coverage fixed only by making the ranges enormous." Full per-ticker, per-year and per-volatility-regime breakdown, plus every rejected alternative, is in results/improvement.json.

Per-ticker: actual price vs. predicted band

Solid line: actual closing price. Bands: predicted range at the time of forecast — muted/hatched for backtest, solid accent for live. A band that misses the line is a miss; width is the cost of getting it right.

Price direction no skill — the negative control

Every credible forecasting study needs a result that's supposed to fail, to prove the evaluation itself isn't rigged to find edges that aren't there. This is that result, scored with the same rigor as the positive findings above — as a probability, not bare accuracy, and against a proper baseline, not a coin flip.

A model with real directional skill would beat simply predicting the historical base rate every time. This one doesn't — its Brier score and log loss are both slightly worse than the base-rate baseline, and its confidence doesn't even rank-order correctly (full calibration curve in results/direction_evaluation.json). That's not a tuning failure to go fix. Markets pricing in almost all easily-available information is the textbook expectation here, and getting exactly that result is what makes the volatility and calibration findings above credible rather than cherry-picked out of a pile of things that were tried.

How this relates to the main study

The main results page asks whether an LSTM beats persistence at next-day close prediction on a fixed 2.5-year held-out period — the answer there is no, and it's the same negative-control pattern this page finds again at a 21-day horizon. This page goes further: alongside the same "does it beat persistence on price" question, it separately asks whether the range around that price prediction is forecastable and honestly calibrated — and there, unlike price direction, the answer is yes. Same tickers, same persistence baseline, same "no lookahead" discipline — and the live evidence keeps accumulating monthly.