Build a baseline NBA points prop model

What you will build

A baseline that predicts a player's points from the average of their own earlier games, scored on later games. It is deliberately simple. A rolling mean tells you what any later model must beat, and it runs on the small statistics samples available today.

Independent ESPN statistics cover NBA, NFL and NHL as limited retained game samples, not full seasons. As of 2026-11-03 those samples were captured after the games were played, so they do not show what was known at game time. Recorded Pick6 lines began around 2026-11-03, so this tutorial does not backtest against past seasons of lines. It cannot, because those lines were not recorded.

export DRAFTKINGS_API_KEY=...
python3 nba-points-baseline.py --season 2027 --min-games 5 --max-pages 5

Pull NBA box scores

NBA seasons use their ending year: season=2027 is 2026-27 and season=2026 is 2025-26. Rows are JSON only, with limit up to 200 and cursor pagination through pagination.next. Use completed=true so in-progress games stay out.

curl --fail-with-body "https://draftkingsapi.com/api/v1/stats/player-games?sport=NBA&season=2027&completed=true&limit=200" \
  -H "Authorization: Bearer $DRAFTKINGS_API_KEY"

{
  "data": {
    "athlete_id": "ESPN_ID",
    "name": "Player Name",
    "category": "boxscore",
    "did_not_play": false,
    "stats": [
      { "key": "minutes", "label": "MIN", "scope": "game", "display": "14", "numeric": 14 },
      { "key": "points", "label": "PTS", "scope": "game", "display": "3", "numeric": 3 }
    ]
  },
  "game": { "starts_at": "START_TIME", "completed": true, "event_id": "ESPN_EVENT" }
}

Check available and coverage in the response. An empty matching page returns available: false, which means no stored rows and does not mean the games did not happen.

Extract the target

For NBA rows, category is boxscore and the target is the stat with key: "points". Read its numeric value and require scope to be game. Skip rows where did_not_play is true or numeric is null rather than scoring them as zero.

def points(row):
    if row["data"].get("did_not_play"):
        return None
    for stat in row["data"]["stats"]:
        if stat["key"] == "points" and stat["scope"] == "game":
            return stat["numeric"]  # None when not a scalar
    return None

Order by game time, then roll

Pagination order follows a cursor built from event and athlete IDs, not the calendar. Sort each athlete's games by game.starts_at before computing anything. The prediction for a game uses only earlier games.

from collections import defaultdict
history = defaultdict(list)
rows.sort(key=lambda r: r["game"]["starts_at"])
for row in rows:
    value = points(row)
    if value is None:
        continue
    earlier = history[row["data"]["athlete_id"]]
    prediction = sum(earlier[-5:]) / len(earlier[-5:]) if len(earlier) >= MIN_GAMES else None
    # record (prediction, value) here, before appending this game
    earlier.append(value)

Appending after predicting is what prevents leakage from the current game. Athletes with fewer than --min-games earlier games get no prediction, and the report counts them.

Score it on later games

Use a date cutoff. Fit nothing on games after it, and report mean absolute error on games after it. MAE reads in points, so it is easy to explain. Report the sample size next to it every time.

• Print the number of athletes and games scored, and how many were skipped and why.
• A small test window gives a noisy number. Do not quote it as model accuracy in general.
• Compare against a trivial baseline such as the player's previous game. A model that cannot beat it has not added anything.

The Python walkthrough covers caching, the cutoff and each piece of the script.

Compare to recorded Pick6 targets

Comparison needs a target you recorded before the game. A Pick6 row read today says nothing about a target last week. Record NBA Points rows, then compare after games finish. Fresh NBA rows depend on the slate, and the table on status shows capture age.

python3 pick6-board.py --sport NBA --stat Points --limit 25 --max-pages 2

• Rows are market rows. One athlete can have several targets, so choose a rule, such as a row with both More and Less, and state it.
• There is no verified mapping from Pick6 entity_id to ESPN athlete_id. If you match on name, refuse ambiguous matches and report the rest.
• Pick6 rows carry standings multipliers such as 1.5x. A standings multiplier is not decimal odds, a probability or an established entry payout.
• A difference between your prediction and a target is a number to inspect. It is not a recommendation, and a small sample cannot show an edge.

Check the recorded data

• Measured coverage report
• Sources and status
• Recorded Pick6 props

Related resources