Data field guide / Accuracy
What does forecast accuracy tell you?
Read average error alongside forecast ranges and sample size. Follow the same fictional models from one forecast to their summary scores.
The forecasts and outcomes below use made-up values. They do not measure any provider’s performance.
Read error and ranges together.
These three invented models use the same two made-up outcomes and an assumed lead time of 1,008 blocks. A third target has no outcome. None represents Doefin or another actual estimator.
One synthetic forecast · Example model A
Start with block 911,232.
- Forecast
- +1.12%
- Outcome
- +1.84%
- Absolute error
- 0.72 pp
Absolute error is the distance between the forecast and outcome, ignoring direction. It uses percentage points (pp), because it compares two percentage changes.
The outcome sits inside A’s range of −1.90% to +4.10%, so this row counts as covered. The range is 6.00 pp wide.
Then average the scored rows.
A’s two absolute errors are 1.32 pp and 0.72 pp. Their average is its mean absolute error: 1.02 pp. Each scored pair has equal weight; the average hides direction and individual misses.
Coverage is the share of scored outcomes inside their ranges. Mean width averages those ranges’ widths. The count shows how many pairs were scored and how many outcomes were missing.
| Fictional model | Scored / missing | Mean absolute error (pp) | Range coverage (%) | Mean width (pp) |
|---|---|---|---|---|
| Example model A | 2 / 1 | 1.02 | 100% | 6.45 |
| Example model B | 2 / 1 | 0.70 | 50% | 1.50 |
| Example baseline C | 2 / 1 | 1.95 | 100% | 10.00 |
Compare the trade-off.
B’s point forecasts were closer on average: 0.70 pp error against A’s 1.02 pp. Its narrower ranges covered fewer outcomes.
A covered 100% with a mean width of 6.45 pp. B covered 50% with 1.50 pp. Both scored 2 outcomes and excluded 1 missing outcome.
A wide range can include more outcomes while telling you less. A narrow range is useful only if it includes outcomes often enough for its stated confidence level.
These two invented outcomes establish neither a better model nor a calibrated confidence level. Testing calibration needs many comparable recorded forecasts.
Optional: inspect every synthetic forecast and outcome
Example model A
- Block 909,216 · synthetic example
Forecast −0.74%; range −4.20% to +2.70%.
Outcome −2.06%.
Input height 908,208; synthetic lead 1,008 blocks. - Block 911,232 · synthetic example
Forecast +1.12%; range −1.90% to +4.10%.
Outcome +1.84%.
Input height 910,224; synthetic lead 1,008 blocks. - Block 913,248 · synthetic example
Forecast +2.31%; range −2.40% to +7.00%.
Outcome not supplied; excluded.
Input height 912,240; synthetic lead 1,008 blocks.
Example model B
- Block 909,216 · synthetic example
Forecast −2.50%; range −3.00% to −1.00%.
Outcome −2.06%.
Input height 908,208; synthetic lead 1,008 blocks. - Block 911,232 · synthetic example
Forecast +2.80%; range +2.00% to +3.00%.
Outcome +1.84%.
Input height 910,224; synthetic lead 1,008 blocks. - Block 913,248 · synthetic example
Forecast +1.50%; range +0.00% to +3.00%.
Outcome not supplied; excluded.
Input height 912,240; synthetic lead 1,008 blocks.
Example baseline C
- Block 909,216 · synthetic example
Forecast +0.00%; range −5.00% to +5.00%.
Outcome −2.06%.
Input height 908,208; synthetic lead 1,008 blocks. - Block 911,232 · synthetic example
Forecast +0.00%; range −5.00% to +5.00%.
Outcome +1.84%.
Input height 910,224; synthetic lead 1,008 blocks. - Block 913,248 · synthetic example
Forecast +0.00%; range −5.00% to +5.00%.
Outcome not supplied; excluded.
Input height 912,240; synthetic lead 1,008 blocks.
These are the complete inputs to this exercise. There are no captured source forecasts or published record pages for these examples.
What a measured comparison needs.
The site has no measured provider comparison: 0 observed adjustment periods scored. Provider errors and range coverage are unavailable.
Historical network observations record what happened. They do not supply the forecasts needed to calculate forecast errors.
Compare the same target and outcome definition at similar lead times. Later forecasts can use information unavailable to earlier ones.
Keep the original timestamped forecasts, source inputs and method versions. State the evaluated dates, scored count, missing forecasts and exclusion rules, including large misses.
A small or selected sample may not represent future conditions. Measured errors cannot guarantee the next forecast, a trading result or production at your fleet.
Observed network research · v1.0.0
Download and cite the dataset.
Source: mempool.space, captured 2026-10-22.
