Position sizing in Polymarket backtests

• Measured 3-day depth ranges from 2,044 (zec) to 57,147 (btc); position size must scale with it
• max_trades_per_market caps how many trades one market contributes to a sample
• The budget parameter bounds total simulated capital (5 to 10,000)
• Full L2 depth in the archive lets you size an order to a fraction of resting liquidity

A fixed contract size is the wrong abstraction because depth varies 28x across assets. A size that is 1% of Bitcoin's top-of-book depth is over half of Zcash's. Sizing as a fraction of measured depth keeps the market impact comparable across assets.

The backtest exposes top-of-book levels per snapshot, so you can set size to a target fraction - say 10% of resting depth on each side. That keeps slippage assumptions consistent and prevents a thin asset from producing fantasy fills.

One market can be entered several times within its window, but treating each as independent overstates the effective sample size. max_trades_per_market caps the trades drawn from a single market, which keeps the sample honest when markets are highly correlated within a window.

On a 5-minute market with a few entries, the cap is usually not binding. On longer windows where a strategy might re-enter repeatedly, it prevents one market from dominating the statistics.

The budget parameter (5 to 10,000) bounds total simulated capital, which matters because a strategy's edge can only be realized at the size the budget allows. A strategy that is profitable you cannot scale is only conditionally useful.

Set the budget to something realistic for your account and re-run. If the measured edge collapses when you raise the budget - because larger orders walk the book - the strategy has a capacity limit, which is a crucial fact to know before deploying it.

On binaries, oversizing is fatal: a run of losses at too large a fraction of capital can end the account before the edge appears. Size so the worst realistic streak does not approach ruin, and use the backtest's max_drawdown as the input to that calculation rather than hoping.

Kelly-style sizing is a useful upper bound, but it assumes the win probability is known exactly - which in a backtest it never is. Halve the Kelly fraction as a starting point and let the drawdown measurement adjust it downward.

# Size an order as a fraction of resting top-of-book depth.
def size_for(depth, fraction=0.10, cap=None):
    qty = depth * fraction
    if cap is not None:
        qty = min(qty, cap)
    return qty

depths = {"btc": 57147, "eth": 26844, "sol": 18132,
          "xrp": 16649, "doge": 12054, "bnb": 11710,
          "hype": 5390, "zec": 2044}
for coin, d in depths.items():
    print(coin, int(size_for(d)))

The Starter plan is free, includes order books, prices, and metrics at 250ms resolution, with 3 days of history, 60 requests/min, 1,000/day, 1 free AI backtest, and 3 strategy backtests. No credit card required.

Paid data windows from $19/mo extend history to 30–120 days at higher throughput. Backtest AI add-on is +$19/mo or standalone at $29/mo.

• Fill model
• Sample size
• Backtest AI