How to read a forecast without fooling yourself

A single price target sounds confident and is almost always wrong. A good forecast gives you a range and a probability, then lets you check later whether those probabilities were honest. Here is how to read one.

Point forecasts versus probabilistic forecasts

“BTC will be at $X next month” is a point forecast. It gives no sense of uncertainty and cannot be fairly evaluated: being off by 1% and by 30% both count as wrong. A probabilistic forecast instead says something like “there is an 80% chance the price will be between A and B”. That statement can be tested. Over many forecasts, about 80% of outcomes should land inside the 80% ranges.

What DUKOIN’s forecast shows

For each supported asset and horizon (24 hours, 7 days and 30 days), the forecast card shows:

• Median: the middle of the forecast distribution. Under our current model it equals the current price, because the model assumes no drift.
• 50%, 80% and 90% intervals: ranges that the model expects to contain the price at the end of the horizon with those probabilities.
• Threshold probabilities: the model’s probability that price ends the horizon at least 5%, 10% or 20% above or below today’s level.
• Direction: a “probability up” figure only if a model has passed a walk-forward test of beating a 50/50 baseline by a statistically distinguishable margin. Otherwise you will see “No demonstrated directional edge”.

A worked example

This is a hypothetical illustration, not a forecast for any real asset. Suppose a token trades at $100 and the model’s 7-day volatility is 10% (about 3.8% per day, scaled by √7).

And the threshold probabilities for the end of the 7-day horizon:

Notice two things. The range is slightly wider on the upside in dollar terms, because prices are modeled in log terms and cannot fall below zero. And the upside and downside probabilities differ for the same reason, not because the model expects a rise. The model is describing the size of likely moves, not their direction.

How to use the numbers

• Sizing. If the 90% range spans a 15% fall, ask whether you could live with that outcome on your current position size.
• Setting alerts. Levels near the edge of the 80% range are moves that the model considers unusual but plausible.
• Comparing assets. Two assets with similar Research Scores can have very different ranges. Wider ranges mean more risk per dollar invested.
• Reality checks. If you are expecting a 40% rise in a week and the model puts a 20% rise at a few percent, you need a strong, specific reason.

Calibration: are the probabilities honest?

A forecaster is calibrated when events it calls 20% likely happen about 20% of the time. DUKOIN tests this in public on the Model performance page, using measures you can learn to read:

• Interval coverage: the share of outcomes that landed inside each interval, compared with its nominal level. An 80% interval with 65% coverage is too narrow.
• Brier score: the mean squared difference between forecast probabilities and outcomes (1 if the event happened, 0 if not). Lower is better; always forecasting 50% scores 0.25.
• Log loss: punishes confident wrong forecasts much more heavily than the Brier score.
• Calibration error: the average gap between forecast probabilities and observed frequencies across probability buckets.

Until at least 30 forecasts have matured for a horizon, the page shows “Insufficient evidence for a reliable calibrated forecast”. Small samples can look great or terrible by chance.

Why the record cannot be rewritten

Every DUKOIN forecast is written to an append-only ledger when it is created, before the outcome is known, and is evaluated automatically when its horizon matures. Records are never edited or deleted after they are issued. This is what makes the track record meaningful: you see the misses as well as the hits.

Why horizons behave differently

Under the model, uncertainty grows with the square root of time. A 30-day range is not 30 times wider than a 1-day range; it is about √30, or roughly 5.5 times wider. That is why the 24-hour card can look calm while the 30-day card shows a wide spread. Longer horizons also leave more room for regime changes the model cannot anticipate: a new trend, a crisis, or a sudden shift in volatility. Treat the 30-day forecast as the least certain of the three, even though its ranges are already the widest.

Volatility estimates also update daily. After a quiet stretch, ranges narrow; after a violent day, they widen immediately. A sudden widening is useful information in itself: the market has become less predictable.

Common mistakes

1. Treating the median as a target. It is the midpoint of a wide distribution.
2. Judging by one outcome. A 10% event will happen one time in ten. That alone does not mean the model failed.
3. Ignoring fat tails. Crypto has more extreme moves than a log-normal model assumes. Expect the 90% interval to be breached somewhat more often during crises.
4. Confusing volatility with direction. A high chance of a large move is not a prediction that price will rise.