Skip to content
EdgeLedger
0x0000…0096demo

You're viewing demo data.

Risk management for predictive market trading

Bankroll sizing, Kelly intuition without the prescription, correlated bets, settlement risk, and platform-specific risks — the things that turn +EV traders into break-even or worse.

·9 min read·Compliance reviewed

EV math describes the expected payoff under a probability estimate; risk management describes the variance and operational failure modes around that estimate. This guide covers the four risks that consistently turn +EV traders into break-even or worse — sizing, correlation, settlement, and platform — and how to think about each without reaching for prescriptions.

If you haven't read the EV groundwork yet, start with Reading prices and expected value and Liquidity, slippage, and order book vs AMM.

Bankroll sizing

The most useful frame is fraction-of-bankroll rather than fixed dollar amounts. Same percentage on every trade scales with your wins and losses; a flat $100 doesn't.

Three observations that set up everything else:

  • Variance compounds. A streak of losses doesn't just hurt the dollar amount — it shrinks the next trade's size too if you're betting a fraction. That's the good property: bad runs taper your exposure automatically.
  • Asymmetry of percent moves. Losing 50% of your bankroll takes 100% to recover, not 50%. Big drawdowns aren't symmetric with the gains it took to get there.
  • The math doesn't care how confident you feel. Confidence is not edge. If your historical hit rate doesn't back up your in-the-moment certainty, your sizing should reflect the history, not the certainty.
Four bettors, same edge, same trade sequence. Higher sizing = higher peaks AND deeper drawdowns. The 30% line nearly went to zero before recovering.

The diagram shows what happens when the same trader (60% hit rate at even odds — comfortably +EV) sizes at 1%, 5%, 15%, and 30% of bankroll per trade across 80 trades. Same edge, same sequence. The 30% line nearly busts before recovering. A few more losses during that drawdown and there's no recovery.

This is why "go big when you're confident" is a story rather than a strategy.

Kelly intuition (without prescribing)

The Kelly criterion is the sizing formula that maximises long-run growth rate of a bankroll given a known edge. For a binary contract that pays $1 at probability p (your belief) when you pay c (the price), full Kelly says to risk a fraction f = (p − c) / (1 − c) of bankroll per trade.

Two intuitions worth carrying away:

  • Edge over price = Kelly fraction. A 10-point edge at a 40¢ price is a 17% fraction. A 1-point edge at a 40¢ price is a 1.7% fraction. Kelly scales sizing with edge, which is the right shape — the question is how much of Kelly to actually use.
  • Full Kelly is optimal long-run, brutal short-run. Maximising growth rate means accepting drawdowns that most traders (correctly) don't want to live through. Full Kelly with a known edge expects ~50% drawdowns at some point during a long run.

The standard practical adjustment is fractional Kelly — size at 1/4 to 1/2 of full Kelly. This trades a little growth for a lot of comfort. The smaller fraction also serves as a buffer against the most common failure mode: overestimating your edge.

This guide deliberately avoids saying "size at X% per trade." The right fraction depends on your edge (which you should audit, not assume), your risk tolerance (which is not the same as your risk capacity), and how much variance you can sit through without bailing on a strategy that's actually working.

Correlated bets

Five different markets aren't five different bets if their outcomes ride on the same underlying driver. A common failure mode: a trader holds five "different" positions that all move together when one news cycle hits.

What looks like diversification can be five expressions of the same underlying view.

How to spot correlation in your own positions:

  • Imagine the headline that moves all of them. If you can write one news story that flips three or more positions, those positions aren't independent.
  • Look at common drivers. Macro markets (rates, recession, inflation) move together when the underlying scenario shifts. Politics markets cluster around election outcomes. Sports markets cluster around the same team's injury report.
  • Treat correlated positions as one larger position for sizing. If three markets move together, sizing each at 5% of bankroll is functionally a 15% position on the underlying driver — even if no single market is more than 5%.

This is also why a fully matched cross-venue YES/NO pair is a special case: the two legs are intended to be anti-correlated by construction. It can still carry fees, timing, eligibility, and rule-mismatch risk.

Settlement risk

Even when the trade is +EV and the size is reasonable, settlement isn't always certain in the way you imagined. The four buckets of settlement risk to think about:

  • Resolution-rule risk. The outcome you bet on isn't the outcome the rules pin to. See How prediction markets work and the per-platform reading-the-market guides (Polymarket, Kalshi).
  • Source-of-truth risk. The official source publishes a result you didn't expect, retracts a story, or fails to publish at all. The market resolves on what the source says, not what you understood the source to say.
  • Dispute risk. Disputes can drag a "clean" market sideways — Polymarket's UMA disputes can run days; Kalshi's centralized reviews can pause settlement. Capital is locked in the meantime.
  • Counterparty / venue risk. A custodial venue can pause withdrawals, freeze accounts during compliance review, or in the worst case fail. An on-chain venue's smart contracts can carry bugs even after audits.

The risk-management point: EV math assumes the rules pay out as modeled. Settlement surprises are one reason to cap single-market exposure in any hypothetical sizing model.

Platform-specific risks

Beyond the cross-cutting risks above, each platform has shape issues worth being concrete about.

Polymarket

  • On-chain custody on Polygon. A wallet compromise is a total loss. Treat the wallet like the bank account it is — see the security basics in the getting-started guide.
  • Wrong-network sends. Assets sent on the wrong network can be permanently stranded.
  • Dispute / oracle risk. UMA is rare-but-real. A determined disputer with a counter-bond can escalate even an "obvious" market to a token-holder vote.
  • Regional eligibility shifts. Polymarket's available jurisdictions can change; what was eligible last year might not be this year.

Kalshi

  • Custodial counterparty. Funds sit with Kalshi. A pause (compliance review, ACH dispute, legal action) means you can't withdraw until it lifts.
  • Funding holds and reversals. Bank, card, crypto, and withdrawal rails can have holds, limits, review periods, or reversals depending on the account and provider.
  • Account-credential security. Kalshi is custodial, so the security boundary is your email + 2FA + password. Lose any of those without backups and recovery is a multi-day support ticket at best.
  • Fee-policy changes. Trading fees, maker/taker treatment, and funding or withdrawal fees can change — a net calculation should use the current schedule.

Cross-venue (when you trade both)

  • Capital locked across two custodians. Even if both legs resolve cleanly, you wait on two separate withdrawal rails to recycle the same dollar.
  • Eligibility mismatches. You may be eligible for one venue but not the other; "the same trade on the other venue" might not be a real option for you.

Putting it together

The shortest-form risk-management process for a +EV trader:

  1. Track your forecasts. A calibration audit (how often did things you said were "70% likely" actually happen?) is the single most useful number you can compute about your own trading.
  2. Size in fractions, not dollars. Same fraction every trade, adjusted only for your edge. Cap the fraction at well below full Kelly.
  3. Treat correlated positions as one. Sum exposures across markets that share a driver, and size against that sum.
  4. Read the rules before modeling size. Resolution-rule surprises change the payoff distribution.
  5. Cap single-market exposure in the model. Confidence is not the same as settlement certainty.
  6. Track venue concentration. Capital held on any venue can be affected by account review, withdrawal holds, provider issues, or legal constraints.

EdgeLedger surfaces realized P&L on closed trades and a portfolio P&L curve — those are the inputs to the calibration audit in step 1. The other steps are decisions, not features.

Where to go next