{"id":2536,"date":"2026-03-25T10:00:00","date_gmt":"2026-03-25T10:00:00","guid":{"rendered":"https:\/\/oddspapi.io\/blog\/?p=2536"},"modified":"2026-10-01T13:20:55","modified_gmt":"2026-10-01T13:20:55","slug":"polymarket-arbitrage-local-bookmakers","status":"publish","type":"post","link":"https:\/\/oddspapi.io\/blog\/polymarket-arbitrage-local-bookmakers\/","title":{"rendered":"Polymarket Arbitrage: A 0.44% Median Edge on $4,023 of Depth"},"content":{"rendered":"

Polymarket arbitrage means you back one side of a market on Polymarket and the other side at a sportsbook, at prices whose implied probabilities add up to less than 100%. On 1 October 2026 at 13:10 UTC we pulled 28 full boards and found 2,938 such pairs. After six screens and a dedupe, 42 distinct arbs were left. The median edge was 0.44%, the best was 1.53%, and the median Polymarket rung behind them held $4,023.<\/p>\n

The rest were placeholder prices, stale quotes, suspended books, off-market lines and skins that copy one feed. A scanner that skips those screens alerts on a 29% “arb” against a $1.15 order. This guide builds the scanner with the screens in place, in Python, against one API that carries Polymarket and 350+ sportsbooks in the same payload.<\/p>\n

Where Polymarket Prices Today<\/h2>\n

We sampled 75 upcoming fixtures across seven sports on 1 October 2026 (kick-offs 1 to 3 October) and asked each one for bookmakers=polymarket<\/code>. Polymarket priced 70 of them: UEFA Nations League soccer, Euroleague basketball, WTA Beijing tennis, MLB, the NFL and college football, NHL and UFC 332.<\/p>\n

Polymarket prices the main markets tight and the wings wide. On the winner market its median overround was 1.0% in basketball, tennis, American football, ice hockey and MMA. Soccer full-time over\/under ran at 0.98%. On NFL and college quarter-by-quarter handicaps and totals the median was 96% to 98%. Those quarter quotes are placeholder rungs with a few dollars behind them, and they produce the biggest fake arbs on the board.<\/p>\n

\n\n\n\n\n\n\n\n\n\n
Factor<\/th>\nPolymarket<\/th>\nSportsbook<\/th>\n<\/tr>\n<\/thead>\n
Price source<\/td>\nOrder book<\/td>\nTrading desk<\/td>\n<\/tr>\n
Winner-market overround (1 Oct 2026)<\/td>\n1.0% median<\/td>\nPinnacle 3.86%, bet365 7.12% on Hapoel Tel-Aviv v Real Madrid<\/td>\n<\/tr>\n
Depth on each price<\/td>\nlimit<\/code> in USD on every rung<\/td>\nnull<\/code> on most books (Pinnacle and exchanges publish it)<\/td>\n<\/tr>\n
Format in OddsPapi<\/td>\nDecimal odds (share price already converted)<\/td>\nDecimal, American, fractional<\/td>\n<\/tr>\n
Separate US feed<\/td>\npolymarket.us<\/code>, its own prices and depth<\/td>\nn\/a<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

polymarket<\/code> and polymarket.us<\/code> are two feeds. They quote different prices and carry different depth. On the worked example below, polymarket.us<\/code> showed 2.50 \/ 1.639 with a limit<\/code> of 0.0 on both sides, while polymarket<\/code> showed 2.439 \/ 1.667 with $4,325 on the home side. The scanner reads polymarket<\/code> and drops any counter-leg whose limit<\/code> is 0.<\/p>\n

Old Way vs OddsPapi<\/h2>\n
\n\n\n\n\n\n\n\n\n\n
Task<\/th>\nPolymarket API + sportsbook scrapers<\/th>\nOddsPapi<\/th>\n<\/tr>\n<\/thead>\n
Get both sides of the arb<\/td>\nGamma and CLOB calls, plus one scraper per book<\/td>\nOne \/odds<\/code> call: up to 290 books per fixture in our sample<\/td>\n<\/tr>\n
Match the market<\/td>\nMap Polymarket slugs to each book’s own IDs<\/td>\nSame marketId<\/code> on every book, same line, same settlement clock<\/td>\n<\/tr>\n
Convert prices<\/td>\nShare price to decimal by hand<\/td>\nDecimal already<\/td>\n<\/tr>\n
Read depth<\/td>\nSeparate order-book call<\/td>\nlimit<\/code> and exchangeMeta<\/code> ladder inline<\/td>\n<\/tr>\n
Backtest<\/td>\nCollect your own history<\/td>\nFree \/historical-odds<\/code>, Polymarket included<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Step 1: Authenticate<\/h2>\n

The key goes in the query string as apiKey<\/code>. The helper retries on a 429, reads the retryMs<\/code> the server sends, and treats a 404 on \/fixtures<\/code> as an empty window.<\/p>\n

import os\nimport time\nimport statistics\nfrom collections import Counter, defaultdict\nfrom datetime import datetime, timedelta, timezone\n\nimport requests\n\nAPI_KEY = os.environ.get(\"ODDSPAPI_API_KEY\", \"YOUR_API_KEY\")\nBASE = \"https:\/\/api.oddspapi.io\/v4\"\n\n\ndef get(path, **params):\n    \"\"\"GET with the key as a query parameter, a 429 retry and a status check.\"\"\"\n    for _ in range(5):\n        r = requests.get(f\"{BASE}{path}\", params={\"apiKey\": API_KEY, **params}, timeout=120)\n        if r.status_code == 429:\n            time.sleep(r.json().get(\"error\", {}).get(\"retryMs\", 2000) \/ 1000 + 0.5)\n            continue\n        if r.status_code == 404 and path == \"\/fixtures\":\n            return []  # empty window, not a failure\n        body = r.json()\n        if r.status_code != 200 or (isinstance(body, dict) and \"error\" in body):\n            raise RuntimeError(f\"{path} {r.status_code}: {body}\")\n        return body\n    raise RuntimeError(f\"{path}: still rate-limited after 5 tries\")<\/code><\/pre>\n
Step 2: Find Fixtures Polymarket Prices<\/h2>\n
A probe with bookmakers=polymarket<\/code> returns a small payload and tells you if Polymarket is on the fixture. Pass tournamentId<\/code> to keep the fixture list short. The free plan gives you 250 requests a month, and each probe counts, so probe one tournament at a time.<\/p>\n
def polymarket_fixtures(sport_id, days=3, tournament_id=None):\n    \"\"\"Fixtures in the next `days` days that Polymarket prices (one small call each).\"\"\"\n    now = datetime.now(timezone.utc)\n    params = {\"sportId\": sport_id,\n              \"from\": now.strftime(\"%Y-%m-%d\"),\n              \"to\": (now + timedelta(days=days + 1)).strftime(\"%Y-%m-%d\")}\n    if tournament_id:\n        params[\"tournamentId\"] = tournament_id\n    hits = []\n    for f in get(\"\/fixtures\", **params):\n        if not f.get(\"hasOdds\") or f[\"startTime\"] < now.isoformat():\n            continue\n        time.sleep(1.0)\n        probe = get(\"\/odds\", fixtureId=f[\"fixtureId\"], bookmakers=\"polymarket\")\n        if \"polymarket\" in (probe.get(\"bookmakerOdds\") or {}):\n            hits.append(f)\n    return hits<\/code><\/pre>\n
Step 3: Flatten the Board<\/h2>\n
One \/odds<\/code> call without a bookmakers<\/code> filter returns every book on the fixture. The NFL board for Cleveland Browns v Pittsburgh Steelers came back at 290 books and 73.9 MB. flatten()<\/code> keeps the moneyline-style fields you need and records each book’s newest changedAt<\/code>, which the freshness screen uses. It also drops the internal test feeds (pinnacle+5<\/code>, pinnacle+30<\/code>, demo<\/code>) that ship in live payloads.<\/p>\n
def ts(s):\n    return datetime.fromisoformat(s.replace(\"Z\", \"+00:00\"))\n\n\ndef flatten(board):\n    \"\"\"{slug: {\"live\": bool, \"newest\": datetime, \"m\": {marketId: {outcomeId: price dict}}}}\"\"\"\n    books = {}\n    for slug, b in (board.get(\"bookmakerOdds\") or {}).items():\n        if slug.startswith(\"pinnacle+\") or slug == \"demo\":\n            continue  # internal test feeds\n        markets, newest = {}, None\n        for mid, m in b.get(\"markets\", {}).items():\n            for oid, o in m.get(\"outcomes\", {}).items():\n                p = o.get(\"players\", {}).get(\"0\")\n                if not p:\n                    continue\n                if p.get(\"changedAt\"):\n                    c = ts(p[\"changedAt\"])\n                    newest = c if newest is None or c > newest else newest\n                if p.get(\"price\"):\n                    markets.setdefault(mid, {})[oid] = p\n        books[slug] = {\"live\": b.get(\"bookmakerIsActive\") and not b.get(\"suspended\"),\n                       \"newest\": newest, \"m\": markets}\n    return books\n\n\ndef overround(pair):\n    return sum(1 \/ p[\"price\"] for p in pair.values()) - 1<\/code><\/pre>\n
Step 4: Screen Every Two-Way Market<\/h2>\n
The scanner compares Polymarket with each book on every two-way market that both price under the same marketId<\/code>. The same ID means the same line and the same settlement clock, so you never pair a regulation-time price with an overtime-inclusive one. Then it applies six screens and a dedupe, and counts what survives each one.<\/p>\n
def scan_board(board, min_depth=500, max_pm_margin=0.05, band=0.10, max_age_h=24):\n    \"\"\"Every Polymarket-vs-book two-way arb on one board, with a count per screen.\"\"\"\n    books = flatten(board)\n    pm = books.get(\"polymarket\")\n    funnel, found = Counter(), {}\n    if not pm:\n        return funnel, []\n    for mid, pm_pair in pm[\"m\"].items():\n        if len(pm_pair) != 2:\n            continue  # two-way markets only\n        # board median per outcome, on distinct active prices\n        median = {oid: statistics.median({b[\"m\"][mid][oid][\"price\"] for b in books.values()\n                                          if oid in b[\"m\"].get(mid, {}) and b[\"m\"][mid][oid].get(\"active\")}\n                                         or {0})\n                  for oid in pm_pair}\n        for slug, book in books.items():\n            pair = book[\"m\"].get(mid, {})\n            if slug == \"polymarket\" or set(pair) != set(pm_pair):\n                continue  # same market ID = same line and same settlement clock\n            for pm_leg in pm_pair:\n                bk_leg = next(o for o in pm_pair if o != pm_leg)\n                p, q = pm_pair[pm_leg], pair[bk_leg]\n                edge = 1 - (1 \/ p[\"price\"] + 1 \/ q[\"price\"])\n                if edge <= 0:\n                    continue\n                funnel[\"0 raw\"] += 1\n                if not (pm[\"live\"] and book[\"live\"] and p.get(\"active\") and q.get(\"active\")):\n                    continue\n                funnel[\"1 active, not suspended\"] += 1\n                if any(not x.get(\"changedAt\") or src[\"newest\"] - ts(x[\"changedAt\"]) > timedelta(hours=max_age_h)\n                       for x, src in ((p, pm), (q, book))):\n                    continue\n                funnel[\"2 fresh\"] += 1\n                if overround(pair) <= 0 or overround(pm_pair) <= 0:\n                    continue  # one venue arbs itself: a broken feed, not an edge\n                funnel[\"3 no self-arb\"] += 1\n                if overround(pm_pair) > max_pm_margin:\n                    continue  # Polymarket placeholder rung\n                funnel[\"4 PM margin <= 5%\"] += 1\n                if any(abs(x[\"price\"] \/ median[o] - 1) > band for x, o in ((p, pm_leg), (q, bk_leg))):\n                    continue  # off-market quote, usually a mapping or stale-line error\n                funnel[\"5 within 10% of board median\"] += 1\n                if (p.get(\"limit\") or 0) < min_depth or q.get(\"limit\") == 0:\n                    continue\n                funnel[\"6 depth\"] += 1\n                key = (mid, pm_leg, p[\"price\"], q[\"price\"])  # dedupe on the price tuple\n                found.setdefault(key, {\"marketId\": mid, \"pm_outcome\": pm_leg, \"pm_price\": p[\"price\"],\n                                       \"pm_limit\": p[\"limit\"], \"book_outcome\": bk_leg,\n                                       \"book_price\": q[\"price\"], \"edge_pct\": round(edge * 100, 2),\n                                       \"books\": []})[\"books\"].append(slug)\n    funnel[\"7 distinct price tuples\"] = len(found)\n    return funnel, sorted(found.values(), key=lambda a: -a[\"edge_pct\"])<\/code><\/pre>\n
This is the funnel on the 28 boards we pulled at 13:10 UTC on 1 October 2026:<\/p>\n
\n\n\n\n\n\n\n\n\n\n\n\n\n
Screen<\/th>\nPairs left<\/th>\nWhat it removes<\/th>\n<\/tr>\n<\/thead>\n
Raw pairs under 100%<\/td>\n2,938<\/td>\nn\/a<\/td>\n<\/tr>\n
Both prices active, neither venue suspended<\/td>\n1,884<\/td>\nParked boards and pulled prices<\/td>\n<\/tr>\n
Each price within 24 h of its book’s newest change<\/td>\n1,064<\/td>\nQuotes the book stopped updating<\/td>\n<\/tr>\n
Neither venue arbs itself<\/td>\n992<\/td>\nBroken two-way pairs on one book<\/td>\n<\/tr>\n
Polymarket overround 5% or less<\/td>\n902<\/td>\nPlaceholder rungs<\/td>\n<\/tr>\n
Both prices within 10% of the board median<\/td>\n214<\/td>\nOff-market quotes and line-mapping errors<\/td>\n<\/tr>\n
$500+ on the Polymarket rung, counter-leg depth above 0<\/td>\n108<\/td>\nOrders you cannot fill<\/td>\n<\/tr>\n
Distinct price tuples<\/td>\n42<\/strong><\/td>\nSkins that copy one feed<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n
The board-median screen does the most work. On the Hapoel Tel-Aviv v Real Madrid board, Duelbits offered Real Madrid at 1.96 against a board median of 1.633. That price was 20% above the median and had not moved since 29 September. Paired with Polymarket’s 2.439 on Hapoel, it shows an 8.0% “arb” that no desk would honour.<\/p>\n
The dedupe step matters too. One 1.22% NCAAF handicap arb showed up eight times, once for each Betano skin quoting the same price. Count it once.<\/p>\n
Step 5: Size the Stakes Against the Depth<\/h2>\n
The equal-payout split puts more money on the shorter price. stakes()<\/code> also caps the total so the Polymarket leg fits inside its top rung.<\/p>\n
def stakes(pm_price, book_price, total, pm_limit):\n    \"\"\"Equal-payout split. Caps the total so the Polymarket leg fits its top rung.\"\"\"\n    s = 1 \/ pm_price + 1 \/ book_price\n    total = min(total, pm_limit * s * pm_price)\n    pm_stake = total \/ (pm_price * s)\n    return {\"polymarket\": round(pm_stake, 2), \"book\": round(total - pm_stake, 2),\n            \"payout\": round(pm_stake * pm_price, 2), \"profit\": round(pm_stake * pm_price - total, 2)}<\/code><\/pre>\n
Run the full scan on the NFL and the Euroleague:<\/p>\n
totals = Counter()\nfor sport_id, tournament_id in [(14, 31), (11, 138)]:  # NFL, Euroleague\n    for f in polymarket_fixtures(sport_id, days=3, tournament_id=tournament_id)[:2]:\n        time.sleep(1.0)\n        board = get(\"\/odds\", fixtureId=f[\"fixtureId\"])  # one call = every book on the fixture\n        funnel, arbs = scan_board(board)\n        totals.update(funnel)\n        print(f\"{f['participant1Name']} v {f['participant2Name']}: {len(arbs)} arbs\")\n        for a in arbs[:3]:\n            print(f\"  +{a['edge_pct']}% market {a['marketId']} | polymarket {a['pm_price']} \"\n                  f\"(${a['pm_limit']:,.0f}) vs {a['books'][0]} {a['book_price']}\")\n            print(\"  \", stakes(a[\"pm_price\"], a[\"book_price\"], 1000, a[\"pm_limit\"]))\nfor stage in sorted(totals):\n    print(stage, totals[stage])<\/code><\/pre>\n
Our run at 13:15 UTC on 1 October 2026 returned 14 distinct arbs across four fixtures. The best one:<\/p>\n
\n\n\n\n\n\n\n\n\n\n
Hapoel Tel-Aviv v Real Madrid (Euroleague, market 111)<\/th>\nHapoel<\/th>\nReal Madrid<\/th>\nOverround<\/th>\n<\/tr>\n<\/thead>\n
polymarket<\/code><\/td>\n2.439<\/strong> ($4,325)<\/td>\n1.667 ($90.78)<\/td>\n0.99%<\/td>\n<\/tr>\n
fanduel<\/code><\/td>\n2.08<\/td>\n1.74<\/strong><\/td>\n5.55%<\/td>\n<\/tr>\n
pinnacle<\/code><\/td>\n2.34<\/td>\n1.636<\/td>\n3.86%<\/td>\n<\/tr>\n
kalshi<\/code><\/td>\n2.50<\/td>\n1.639<\/td>\n1.01%<\/td>\n<\/tr>\n
Board median (252 active books)<\/td>\n2.24<\/td>\n1.633<\/td>\nn\/a<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n
Hapoel on Polymarket at 2.439 plus Real Madrid at FanDuel at 1.74 sums to 98.47%, a 1.53% edge. On $1,000 you stake $416.37 on Polymarket and $583.63 at FanDuel, and either result pays $1,015.52, a $15.52 profit before fees. FanDuel’s price had last changed at 20:53 UTC the night before. Check that the book still offers it before you place the Polymarket leg.<\/p>\n
Step 6: Backtest on Polymarket History<\/h2>\n
Historical odds are free, and they include Polymarket. \/historical-odds<\/code> takes Polymarket only on its own and only with one outcomeId<\/code>. A call with bookmakers=polymarket,pinnacle<\/code> returns 400 INVALID_PARAMETER<\/code>.<\/p>\n
# Polymarket on \/historical-odds: one bookmaker and exactly one outcomeId per call\nhist = get(\"\/historical-odds\", fixtureId=\"id1400003171515810\",  # Bills v Chargers, 27 Sep 2026\n           bookmakers=\"polymarket\", outcomeId=\"141\")\nsnaps = hist[\"bookmakers\"][\"polymarket\"][\"markets\"][\"141\"][\"outcomes\"][\"141\"][\"players\"][\"0\"]\npre = sorted((s for s in snaps if s[\"active\"] and s[\"createdAt\"] < \"2026-09-27T17:00:00\"),\n             key=lambda s: s[\"createdAt\"])\nclose = pre[-1]\nprint(f\"{len(snaps):,} snapshots, {len(pre):,} active before kick-off\")\nprint(f\"closing price {close['price']} with ${close['limit']:,.2f} on the top rung\")\nprint(\"ladder in history:\", close[\"exchangeMeta\"])<\/code><\/pre>\n
On Buffalo Bills v Los Angeles Chargers (27 September 2026) this returns 20,046 snapshots, 14,222 of them active before kick-off, and a closing price of 1.333 with $27,819.28 on the top rung. The exchangeMeta<\/code> ladder is null<\/code> in history, so limit<\/code> is the only depth field you can backtest on. Sleep 4.6 seconds after each historical call.<\/p>\n
Use it to measure:<\/p>\n