Every pending transaction, best-paying on the left. Colour separates what arrived today from what has been waiting over a day.
How the probability is worked out
Measured, not derived. Since 2 September 2026 this node records when every
transaction enters and leaves its mempool, which the chain never stores and
nobody can reconstruct afterwards. Transactions are grouped by effective
fee rate, the package rate miners rank by, so a cheap child counts at the rate
that actually gets it mined; and by how deep the backlog above it was on arrival.
The answer is how often that group confirmed in time. Groups under 500
transactions go unanswered.
Accuracy, out of sample
Refitted nightly without the most recent blocks, then tested on exactly those.
Lower Brier is better, and the bar that matters is beating the long-run average.
Limits
Days of one quiet fee regime, not months, so this says nothing about a fee spike.
Even a very large fee tops out near 94% for the next block: a miner already
working on an older template cannot include you whatever you pay. A queueing model
ran here until 3 September 2026, retired for scoring worse than the long-run
average at every horizon past one block. This page knows fee rates, not your
transaction, so it cannot see replace-by-fee bumps, out-of-band mining, or your
transaction being evicted before it confirms.
Papers
Closest to what runs here:
A hybrid
model combining neural networks and XGBoost,
Confirmation
Time Prediction: A Classification View.
The retired queueing approach:
A Holistic Approach for Bitcoin
Confirmation Times and Optimal Fee Selection,
Predicting the confirmation time of
Bitcoin transactions,
Predicting
confirmation times of Bitcoin transactions, Gundlach and co-authors.