The FC27 live test

The model from the report is continuing into EA Sports FC 27, the season now being played. Each week it makes a call on every card in its universe, and each call is frozen before its outcome can be known. This page publishes a fingerprint of each week's calls before the week resolves, reveals the calls once the week's outcome is in, and scores them.

The two models

Both were locked on 25 September 2026, before any FC27 outcome existed, as fixed files with the SHA-256 fingerprints below. Neither changes during the season.

• The live model is the report's gradient-boosted trees (configuration B), with its settings re-chosen by a selection procedure fixed before it ran, and trained on FC24, FC25 and FC26. Its calls are the headline.
• The benchmark is the report's recipe (configuration B, sized by the report's rule), re-fit once on FC24 and FC25 and frozen; it was never trained on FC26 or FC27.

• Live model (fc27_champion): model file, then spec filee63d7be17745ed6672aae7149c78980ffc16328ff4d19444a6e4b75a6c6c54fbf719833b9d52ed1a3c47d3ded325ae5427db47621f38e60b3cfe937f37960da7
• Benchmark (rung2_lgbm_B): model file, then spec file9b59a57ac8509b281059303434e44f9755643fc146dabd42c5d3629054a6fde66a7e4eaedcb13d03508bfcdf0163f2fa3c6c72471e5c94b7a2272e5912a2400f

Both are compared with the report's baselines: the lifecycle prior (the bar) and buying everything. As in the report, a call is a buy when the predicted 14-day move clears the 5% sales tax.

FC26, final

FC26 closed on 12 September 2026. The benchmark finished its out-of-sample season at:

These replace the report's FC26 figures (MCC 0.202 against the prior's 0.088; +0.117 coins per trade, or +11.7%). The season has since finished, and the benchmark is now scored by the frozen re-fit above rather than the report's June fit. The anchors from June 2026 on, 43% of the hold-out, carry end-of-season returns at a much larger scale, and whether their prices are real is still being checked; without them the trimmed figure is about +9%.

How the ledger works

Every call is written once, when first published, to an append-only table in the lakehouse, stamped by the server's clock. The table refuses any call stamped at or after the moment its outcome can be known, 14 days after the week's decision day, and it is append-only: a call can't be edited or removed without a logged change to the table itself, which, once the week's fingerprint is posted here, the check below would expose.

For each week this page posts a SHA-256 fingerprint of the full call list at the next update after publication, before the week resolves (a fingerprint recorded later is marked), and the list itself once it has resolved. The list is a CSV of both models' calls, one row per model and card: model, item_id (the card's EA id), pred_class (0 down, 1 hold, 2 up), in_trade (true for a buy) and pred_logret (the predicted 14-day log-return, to 6 decimals), sorted by model, then item_id as a number, in UTF-8 with LF line endings. To check a revealed list against its fingerprint:

curl -s https://vinaysankar2004.github.io/futvault-report/fc27/calls/2026-09-17.csv \
  | shasum -a 256

It must print the fingerprint listed below (sha256sum works too); changing any character changes it. The fingerprints are also in commitments.json, and its commit history records when each was posted.

What counts

A call counts if it was published before its outcome could be known, which the ledger enforces. None is dropped for being late, but lateness is disclosed: each week shows its publication lag, the time from 00:00 UTC on its decision day to first publication, and the results are also given for on-time calls only, those first published within 2 days.

The first two weeks were first published at 00:18 UTC on 26 September 2026, late; the ledger was created later that morning and recorded them with that publication time. Later weeks are published by the daily pipeline, expected on the decision day itself.

A call on a card with no recorded price 14 to 17 days later has no outcome and is left out of the results, the rule the report's evaluation used.

The ledger

2 weeks published, none resolved. The next to resolve is 17 Sep, on 1 October 2026. Buy calls are the live model's, then the benchmark's; times are UTC.

The fingerprints (SHA-256), one per publication:

• 17 Sep · calls/2026-09-17.csv, revealed from 1 October 2026 · fingerprint recorded 26 September 2026005b6bb9861284ba5c9137b3926468d3f1047a1ba639b8aad4d439244685831b
• 24 Sep · calls/2026-09-24.csv, revealed from 8 October 2026 · fingerprint recorded 26 September 2026ac08c57dc60d18b17ca2bfd994da8aa1de40a5152f23204b3abb8cdc0552748d

Results

No week has resolved yet. Figures appear from 8 October 2026, after that day's pipeline run, once 2 weeks have resolved; a single week is too few to read.