dataforge-labs/bitcoin-fee-estimator-accuracy

Bitcoin fee estimator accuracy

What five fee estimators recommended, compared with sampled realized block fee percentiles. This is an observational benchmark, not a guarantee that a transaction would have confirmed.

Only one half of this is scarce, and it is worth being precise about which.

Outcomes are freely available. One of these providers will serve you a year of realised per-block fee rates for the asking, so what a block required is not a secret and never was.

The advice is the perishable half. Every provider answers for right now and none serves a history of what it said: the documented history endpoints return 404, and the one that looks like an archive turns out to hold realised rates rather than past forecasts. A recommendation made last month therefore exists only if somebody wrote it down at the time.

That is what this repo is. Not a clever method -- the join is arithmetic anyone could do -- but the half of the inputs that cannot be obtained after the fact.

Contents

name	one row is
e15_fee_estimators	one provider's recommendation at one moment, for one confirmation target
e28_fee_estimator_accuracy	one provider's recommendation compared with a sampled block fee percentile

Reading it

sufficient is a legacy column name: it means only that the recommendation was at least cleared_p10. It does not prove confirmation. overpay_ratio divides the recommendation by that percentile; it is not a measured saving or a counterfactual transaction cost.

target_blocks includes approximate conversions of minute targets and provider priority labels. The legacy join uses target_blocks times 600 seconds, not the next actual N blocks. New E15 rows preserve response-availability timestamps, raw payloads, native targets and units. New E28 rows carry metric_kind=sampled_fee_percentile_proxy and proves_confirmation=false; older partitions predate those columns but have the same limitations.

A separate private E31 panel records sampled real transactions, pending follow-up across runs, package changes, canonical confirmation checks and observation gaps. It does not infer that senders followed a provider's recommendation or broadcast experimental transactions.

What counts as the rate that cleared

cleared_p10 is the 10th percentile fee rate among standalone transactions in the block: those with no unconfirmed parent.

That restriction is doing real work and the number is wrong without it. About 93% of the cheapest-looking transactions in a block are not standalone -- they were admitted on a relative's fee, through a parent or a fee-bumping child. Counting them measures the lowest rate visible in a block rather than the lowest rate that would have worked, and it inflates apparent overpayment by roughly three and a half times. cleared_p10_all_txs carries the unrestricted figure so the gap between the two definitions is visible in the data.

The minimum is not used at all: blocks legitimately contain zero-fee transactions handed straight to a pool, and those measure a private arrangement rather than a market.

Before you build on this

• Five providers, and they are the ones that answer without an API key. That is a selection.
• n_standalone is on every row. A block with few standalone transactions gives a noisier percentile, and low-activity periods produce fewer of them.
• Fee coverage in the underlying panel is partial and varies with load, so the percentile is computed over transactions whose fee we sampled, not over the whole block. It is never imputed.
• A quiet mempool flatters every provider: when almost anything confirms, being sufficient is easy and percentile ratios are large. Read the two columns together, and read them by regime rather than pooled across months.
• lead_seconds records how far ahead of the block each forecast was actually made. Matching is to the newest forecast at or before the target moment, so this varies with sampling cadence.

Partitions are parquet, one file per collection window, under dataset/YYYY/MM/. Every dataset here carries a FIXED 7-day sample WINDOW starting at its own first day of collection, together spanning 2026-08-26 to 2026-09-05, so you can check schema, coverage and quality before asking for more. It does not advance, so there is nothing to gain by re-downloading it. The full history is held privately, available on request.

from huggingface_hub import snapshot_download
import pandas as pd, glob

path = snapshot_download("dataforge-labs/bitcoin-fee-estimator-accuracy", repo_type="dataset",
                         allow_patterns="e28_fee_estimator_accuracy/**")
df = pd.concat(map(pd.read_parquet,
                   glob.glob(f"{path}/e28_fee_estimator_accuracy/**/*.parquet", recursive=True)))

Coverage

e0_run_manifest lists every collection window with its poll counts and failure counts, and is published in full rather than windowed. Gaps between windows are real, cannot be filled in afterwards, and nothing here is interpolated.

License and contact

ODC-BY: use it freely, credit "DataForge (dataforge-labs)". Questions and requests for the full history via the discussions tab.