The volatility projection

Understand what the band drawn beyond the last candle shows: a measured estimate of how far the price could move, not a prediction of direction.

The essentials

Beyond the last candle, the chart can draw a cone: a band that widens progressively and represents the price levels judged probable at the coming horizons. This band is built from the asset's recent volatility — measured live, from the last hundred candles of the asset shown —, then projected using quantiles that instead come from a table measured once and for all on the full history of three reference assets (BTCUSDT, ETHUSDT, SOLUSDT), not recomputed live for each asset: that requires several thousand independent candles, a depth of history not yet available for every listed asset. A per-asset estimate is achievable on the shorter intervals once that data depth is reached — but not on the 1-year range: the available daily history there will remain structurally insufficient, and the measured table will remain its source.

The middle line of the cone is perfectly flat, and that is not an oversight: it stays equal to the last known price, at every horizon. Nobody — not this model, not any other — knows which direction a crypto price will move in over the coming hours or days. What can be estimated, instead, is how far the price could move around its current level. The cone answers the question "how much could the price move?", never the question "up or down?".

The narrow band (50%): half of past moves of this size stayed inside it.
The wide band (80%): four out of five past moves stayed inside it — one in five broke out of it, in either direction.
The farther the horizon, the wider the band: uncertainty about the size of the move grows with elapsed time.

The cone is recalculated every time the range changes: it uses the horizon and quantiles specific to the candle interval shown, not a generic rule that would apply across every timeframe.

What the cone does not tell you: it says nothing about which direction the price will take, only about its probable size of movement. It does not account for upcoming news (regulation, central bank decisions, project-specific announcements) or for cascading liquidations that can suddenly amplify a move. And 20% of the time, the price moves outside the 80% band — that is the expected behavior of a band calibrated at 80%, not a failure of the model.

Going further: the measured calibration

These bands are not set by assumption: they are measured. A validation harness replays the history of three reference assets (BTCUSDT, ETHUSDT, SOLUSDT) and counts, for each time range, the real proportion of times the future price stayed inside each band, over disjoint windows — that is, tests independent of one another. The most recent report is dated 23 August 2026.

1 day: 50% band hit at 49.9%, 80% band at 79.4%, over 2,748 independent tests.
7 days: 50% band hit at 49.8%, 80% band at 79.4%, over 2,331 independent tests.
1 month: 50% band hit at 50.4%, 80% band at 79.5%, over 1,886 independent tests.
3 months: 50% band hit at 51.0%, 80% band at 79.7%, over 585 independent tests.
1 year: 50% band hit at 51.8%, 80% band at 82.1%, but over only 125 independent tests.

The 1-year range does not carry the same standard of proof as the other four. With only 125 aggregated independent tests, against more than 1,800 for the 1-month range, it is measured but far less precise: the validation harness explicitly classifies it in a "weak" regime, as opposed to the "strict" regime of the other four ranges. The figures above for 1 year remain the best estimate available, but with a markedly wider margin of uncertainty than elsewhere.

Expert: the statistical choices behind the cone

Per-candle volatility is measured with the Rogers-Satchell (1991) estimator, which uses each candle's open, high, low, and close. It is preferred over Garman-Klass and Parkinson, two older estimators that both assume zero price drift over the period: they then mistake a pronounced trend for volatility and overestimate it accordingly. A liquid crypto asset is nearly always trending, even mildly, over the windows considered here — Rogers-Satchell is precisely built to stay correct in that case, by remaining independent of drift.

The cone's bounds do not come from a known probability distribution: they are empirical quantiles, measured directly from thousands of historical residuals (the actually observed gap between the price at the horizon and the starting price, normalized by local volatility). This choice is not cosmetic: the study that preceded this work found that a normal distribution produces a 50% band that is 6 percentage points too wide, and that a Student's t distribution with 4 degrees of freedom does even worse — the real distribution of residuals is more peaked at the center than a normal one, while keeping tails close to a normal's. No standard parametric distribution fits both the center and the tails at once.

These residuals are centered on their median before quantiles are taken from them. Without this centering, the historical sample carries the drift of the period it was measured over — the asset broadly rose or fell across the window used — and the cone's bounds would inherit that drift, making the chart imply a direction the product explicitly claims never to predict. Centering is done on the median rather than the mean: it is the median, not the mean, that must land exactly on the last close, since the cone's middle line is by construction that last close.

The validation harness applies two statistical tests to every range/asset pair. The Kupiec (1995) test checks that the observed exception rate matches the expected rate, on average. The Christoffersen (1998) test additionally checks the independence of exceptions over time: a good risk model must not only be wrong at the right average rate, it must also not be wrong twenty times in a row while a market is collapsing, which would make the band dangerously optimistic exactly when it matters most.

Across the fifteen range/asset pairs measured, three fall below the 5% threshold on this independence test — where testing fifteen combinations should, by chance alone, put fewer than one below it. Only one crosses the Bonferroni-corrected threshold (0.05 / 15 ≈ 0.0033), which exists precisely to correct for this number of tests: ETHUSDT on the 7-day range, at p ≈ 0.0008, well below it. Centering the residuals on their median (see above) made this clustering worse, not better. In practical terms for the user: on this asset and this range, when the cone is wrong, it tends to be wrong several times in a row rather than once in isolation — exactly what the Christoffersen test is built to catch, and exactly what would make the band dangerously optimistic during a prolonged market shock on this asset at this horizon. This result is reported here rather than left out, precisely because a cone that claims rigor must also report what does not confirm it perfectly.