FiveThirtyEight's NBA predictions are gone. What to use now.

Switch in one line

nba-538-format.csv holds every game of the 2026-27 season that has Clutch's number (56 so far), rebuilt four times a day. In Python:

import pandas as pd

# was: https://projects.fivethirtyeight.com/nba-model/nba_elo_latest.csv
games = pd.read_csv("https://clutchledger.com/data/nba-538-format.csv")

# code that reads elo_prob1 / elo_prob2 can point at Clutch's columns:
games = games.rename(columns={"clutch_prob1": "elo_prob1", "clutch_prob2": "elo_prob2"})

In R:

games <- read.csv("https://clutchledger.com/data/nba-538-format.csv")

The same games as JSON, with each call's result and seal, are on the open data page.

What's the same

One row per game, with FiveThirtyEight's first columns in their order: date, season as the year the season ends (2027 for 2026-27), neutral as 0 or 1, playoff blank in the regular season, team1 at home and team2 away with FiveThirtyEight's team codes (BRK, CHO, PHO), and score1 and score2 once the game is final.

What's different

The number. It is Clutch's, not Elo or RAPTOR: a model of each team's form, rest, travel and the injury report, blended with the market's price before tip. It sits in clutch_prob1 and clutch_prob2; there are no team-rating columns.

The years. This season only, from opening night. FiveThirtyEight's file ran from 1946 to the 2023 Finals, and the Internet Archive keeps copies of nba_elo.csv and nba_elo_latest.csv for anyone who needs those seasons; Clutch's sealed numbers start this season.

The schedule. A game's row appears once Clutch has a number for it, usually the morning of the game, not weeks ahead.

The record. Every number is sealed with a SHA-256 fingerprint in a public repository before tip, so no row can be changed after the final. How to check one →

How good is it?

This season so far Clutch's calls are 36–16 on 52 graded games, every one sealed before tip. Replayed over two seasons (2024-25 and 2025-26, 2,631 games), each game predicted only from the games before it, Clutch's probabilities were more accurate than ESPN BPI's in both seasons (clearly in 2025-26; in 2024-25 the gap was within the noise) by the accuracy score, which grades the probability itself. Against the market's final price they were level: as good as the market, not better. On winners alone, the three were within 1.3 points of each other each season.

The live record, band by band → · ESPN BPI and the market on the same games →

Other places to look

ESPN's BPI gives a win probability for every game on ESPN's site, and the market's price can be read as one too. Clutch's game pages show both beside its own number, and the Ledger grades all three on the same games every night. The Ledger →