Build an NBA prop model.

Step 1: pull the graded lines to train on

Start from the training set: one flat row per graded NBA line, with the features known before the game and the result. Drop the pushes, keep the rows where the before-game features are filled in, and split features from label.

import os, requests

API = "https://prizepicksapi.com/api/v1"
H = {"Authorization": f"Bearer {os.environ['PRIZEPICKS_API_KEY']}"}

rows = requests.get(f"{API}/training-set", headers=H, params={
    "league": "NBA", "tier": "standard", "limit": 1000,
}).json()["data"]
rows = [r for r in rows if r["result"] != "push"]
rows = [r for r in rows if r["rest_days"] is not None and r["home_away"] is not None]

X = [[r["closing_line"], r["rest_days"], r["home_away"] == "home"] for r in rows]
y = [r["result"] == "over" for r in rows]

Step 2: add recent form from game logs

The line alone is weak. Add the player’s recent box scores as features, read before the game you are scoring. Game logs are full-game box scores; a stat missing from stats was not reported and is not zero, so handle absent keys rather than defaulting them.

logs = requests.get(f"{API}/game-logs", headers=H, params={
    "player": "Jayson Tatum", "league": "NBA", "last": 10,
}).json()["data"]
# Average only the games before the one you are scoring; never the same game.
pts = [g["stats"].get("points") for g in logs if g.get("minutes")]

Step 3: score today’s board

Fetch the current board for the market, build the same features for each line, and compare your estimate with the line. Read observed_at and only act on lines that are still current.

board = requests.get(f"{API}/projections", headers=H, params={
    "league": "NBA", "stat": "Points", "view": "compact", "limit": 1000,
}).json()["data"]

Avoid the two mistakes that ruin a model

• Leakage: minutes, actual_value and result are known only after the game. Use them as labels or as past-game features, never as features of the same game.
• Survivorship: do not train only on lines that still exist today; that drops everything that went unavailable. Use the recorded history as of the decision time.