Data for a PrizePicks prop model.

What one training row looks like

GET /api/v1/results returns one row per line. Use line as the target threshold: it is the last line observed before the start, and closing.gap_seconds says how long before. Treat actual_value and result as labels only.

Pull every graded row with Python

This writes one CSV row per graded line, paging with the numeric cursor. It uses only the standard library, so it runs anywhere.

import csv, json, os, time
from urllib.parse import urlencode
from urllib.request import Request, urlopen

KEY = os.environ["PRIZEPICKS_API_KEY"]

def results(league="NBA"):
    cursor = None
    while True:
        params = {"league": league, "status": "graded", "limit": 200}
        if cursor:
            params["cursor"] = cursor
        req = Request("https://prizepicksapi.com/api/v1/results?" + urlencode(params),
                      headers={"Authorization": "Bearer " + KEY})
        with urlopen(req, timeout=30) as res:
            page = json.load(res)
        yield from page["data"]
        cursor = page["meta"]["next_cursor"]
        if not cursor:
            return
        time.sleep(1.1)  # stays under the Free plan's per-minute limit

with open("training.csv", "w", newline="") as f:
    out = csv.writer(f)
    out.writerow(["date", "player", "stat", "tier", "line", "opening_line", "moved", "actual", "over"])
    for r in results():
        if r["result"] not in ("over", "under"):
            continue  # pushes carry no over or under label
        moved = r["line"] - r["opening_line"] if r["opening_line"] is not None else ""
        out.writerow([r["game_date"], r["player"]["name"], r["market"]["stat"], r["sides"].get("more", {}).get("tier"),
                      r["line"], r["opening_line"], moved, r["actual_value"], int(r["result"] == "over")])

Keep the future out of your features

• Split by date, never at random. Lines for the same game are correlated, and a random split leaks them across folds.
• Build rolling player features only from games before game_date.
• Use line and opening_line, not any value recorded after the start.
• Report sample size per market and tier. Small markets swing hard.
• Drop rows with a large closing.gap_seconds if you act close to the start.

Measure against the base rate first

The table above this guide is the number to beat. A model that predicts the over rate for each market and tier has real skill only if it improves on that baseline out of sample. Check calibration too: a model that says 60% should hit about 60% of the time.

Score the live board

GET /api/v1/projections returns the same player, market and line fields for the current board. Pass view=compact for up to 1,000 rows per page. Run your features for those rows, then keep the line timestamp with each prediction so you can grade it later against /api/v1/results.

curl -G "https://prizepicksapi.com/api/v1/projections" \
  -d league=NBA -d stat=Points -d view=compact -d limit=1000 \
  -H "X-User-Request: Read PrizePicks data for an integration check" -H "Authorization: Bearer $PRIZEPICKS_API_KEY"