DEV Community

Cover image for Polymarket TWAP Pricing: How Market Makers Should Quote Around a Moving Reference
Nagi
Nagi

Posted on

Polymarket TWAP Pricing: How Market Makers Should Quote Around a Moving Reference

#polymarket #twap #pricing #makers

Learn how Polymarket TWAP pricing can be modeled for dynamic quotes using reference divergence, volatility, inventory, and time-to-resolution.

Introduction

A market maker quoting around a TWAP has a different problem from one quoting around a conventional spot price.

A spot price can be treated as a relatively immediate estimate of current fair value. A TWAP is a stateful reference: its current value depends on prices observed throughout an averaging window. When the underlying asset moves quickly, the TWAP can therefore lag the information arriving in the market.

That lag creates an interesting market-making problem. If quotes are centered mechanically on the latest TWAP, they may become stale precisely when informed traders have the strongest reason to trade against them.

Contacts

Nagi writes about Polymarket bots, algorithmic trading, quantitative strategies, Python automation, Web3, and prediction-market infrastructure.

Github: https://github.com/NagiPoly/poly-maker
Telegram: https://t.me/nagi_777x

The useful question is therefore not:

“What is the current TWAP?”

It is:

“Where should the next quote be relative to a TWAP that is still moving toward the underlying market?”

The Core Question

Suppose the reference price is a moving TWAP (T_t), while the market maker observes a faster underlying reference (S_t).

A naive quoting model is:

P_{mid,t}=T_t
Enter fullscreen mode Exit fullscreen mode

with:

Bid=T_t-\frac{s}{2}
Enter fullscreen mode Exit fullscreen mode 
Ask=T_t+\frac{s}{2}
Enter fullscreen mode Exit fullscreen mode

where (s) is the desired bid-ask spread.

The problem is that (T_t) may be lagging (S_t).

A better model introduces a TWAP velocity adjustment:

F_t=T_t+\lambda(S_t-T_t)
Enter fullscreen mode Exit fullscreen mode

where:

  • (T_t) = current TWAP
  • (S_t) = faster external reference
  • (F_t) = estimated fair value for quoting
  • (\lambda) = estimated correction for TWAP lag

This is not a claim that (S_t) is the correct resolution price. It is a hypothesis about where the TWAP is heading.

That distinction matters.

What We Are Analyzing

Consider a short-duration crypto prediction market whose resolution depends on a Chainlink TWAP.

The research dataset should contain:

  • timestamped TWAP observations
  • underlying spot/reference observations
  • Polymarket bid and ask prices
  • order-book depth
  • executed trades
  • time remaining until resolution
  • inventory of the market maker

For research purposes, synthetic data is sufficient for the first experiment.

The important variable is not simply the TWAP itself, but its distance from the faster reference:

D_t=S_t-T_t
Enter fullscreen mode Exit fullscreen mode

and its rate of change:

V_t=\frac{T_t-T_{t-\Delta}}{\Delta}
Enter fullscreen mode Exit fullscreen mode

The first measures lag. The second measures TWAP momentum.

A Better Quote Model

A practical research model can separate three components:

F_t=T_t+\lambda D_t-\gamma I_t
Enter fullscreen mode Exit fullscreen mode

where (I_t) represents inventory exposure and (\gamma) controls inventory skew.

The resulting quotes become:

Bid_t=F_t-\frac{s_t}{2}
Enter fullscreen mode Exit fullscreen mode 
Ask_t=F_t+\frac{s_t}{2}
Enter fullscreen mode Exit fullscreen mode

The interesting part is (s_t).

A static spread ignores changing information risk. Instead:

s_t=s_0+\alpha |V_t|+\beta \sigma_t+\delta R_t
Enter fullscreen mode Exit fullscreen mode

where:

  • (s_0) = base spread
  • (|V_t|) = magnitude of TWAP movement
  • (\sigma_t) = estimated short-term volatility
  • (R_t) = resolution or event risk
  • (\alpha,\beta,\delta) = parameters estimated from historical data

This produces a useful principle:

A rapidly moving TWAP should generally change both the center of the quote and the width of the quote.

Moving only the midpoint can leave the market maker exposed to adverse selection.

The Key Research Insight: TWAP Lag Is State-Dependent

TWAP lag is not constant.

If the underlying price moves smoothly, the TWAP can provide a relatively stable reference. During a sudden move, however, the difference between the underlying price and the TWAP can expand quickly.

This suggests splitting observations into regimes:

  1. Stable: low (D_t), low (V_t)
  2. Trending: persistent (D_t) and directional (V_t)
  3. Shock: rapidly increasing (|D_t|)
  4. Mean-reverting: (D_t) shrinking toward zero

The market maker can then estimate fill outcomes separately for each regime.

This is more informative than calculating one average spread across the entire dataset.

Practical Example

Consider a hypothetical market where:

T_t=0.54
Enter fullscreen mode Exit fullscreen mode

while a faster reference indicates:

S_t=0.58
Enter fullscreen mode Exit fullscreen mode

Suppose historical estimation suggests:

\lambda=0.50
Enter fullscreen mode Exit fullscreen mode

Then:

F_t=0.54+0.50(0.58-0.54)=0.56
Enter fullscreen mode Exit fullscreen mode

Instead of centering quotes around 54%, the model centers them around 56%.

If the required spread is 4 percentage points:

Bid=0.54
Enter fullscreen mode Exit fullscreen mode 
Ask=0.58
Enter fullscreen mode Exit fullscreen mode

The important observation is that the model has not predicted the final outcome. It has estimated the direction and magnitude of reference-price adjustment.

That distinction makes the framework testable without pretending that a TWAP forecast is equivalent to a resolution forecast.

What Can Go Wrong?

The largest risk is treating the faster reference as ground truth.

A Polymarket market can resolve according to a specific reference mechanism. The external spot price may therefore differ from the value ultimately used for resolution.

Other failure modes include:

  • stale TWAP observations
  • incorrect timestamp alignment
  • exchange-price discrepancies
  • latency between reference updates
  • sudden volatility
  • thin order-book liquidity
  • adverse selection
  • inventory accumulation
  • spread assumptions that ignore fees
  • look-ahead bias when estimating (\lambda)
  • fitting one lag coefficient across different market regimes

The most dangerous backtesting mistake is using future TWAP observations when estimating what the market maker could have known at quote time.

Testing Methodology

Separate the research into four stages:

Hypothesis: TWAP lag contains information about the direction of future TWAP movement.

Experiment: Calculate (D_t), (V_t), volatility, and time-to-resolution using historical observations.

Observed result: Measure subsequent TWAP movement after 1, 5, 10, or 30 seconds.

Interpretation: Determine whether the signal remains useful after accounting for spread, latency, and execution constraints.

Walk-forward testing is preferable to randomly shuffling observations because the market is fundamentally time-dependent.

A useful metric is:

E[\Delta T_{t+h}|D_t]
Enter fullscreen mode Exit fullscreen mode

which asks how future TWAP movement varies conditional on the current TWAP/reference divergence.

Advanced Extensions

Experienced researchers can extend the framework in several directions.

1. Kalman filtering: Estimate latent fair value while explicitly modeling noisy observations.

2. Regime detection: Use hidden-state models to distinguish stable markets from shocks.

3. Dynamic spreads: Estimate adverse-selection risk directly from historical fills.

4. Cross-market signals: Compare related Polymarket contracts or external market references.

5. Online estimation: Update (\lambda) continuously rather than assuming a fixed lag relationship.

Key Takeaways

  • A TWAP is a moving reference, not necessarily the fastest estimate of current information.
  • Quote adjustment should consider both TWAP level and TWAP/reference divergence.
  • The same TWAP lag can have different implications in stable and shock regimes.
  • Spread width should respond to information risk, not only volatility.
  • Historical testing must prevent look-ahead and timestamp leakage.
  • Resolution methodology and trading-price methodology should remain separate in the research model.

FAQ

Should market makers quote directly around the TWAP?

Not necessarily. A TWAP can lag a faster-moving reference, so a lag-adjusted fair-value estimate may be more appropriate for research.

Does a higher TWAP mean the prediction probability should immediately increase?

Not automatically. The TWAP is only one input. The relationship between the reference asset and the binary outcome must still be modeled.

What is the most important variable besides TWAP?

The divergence between the TWAP and a faster reference is particularly useful because it measures potential lag.

Should spreads remain constant?

A constant spread is a simple baseline. A research system can instead condition spread width on volatility, TWAP velocity, liquidity, inventory, and time-to-resolution.

How should this strategy be validated?

Use historical replay, walk-forward estimation, synthetic stress tests, and separate evaluation of fair-value prediction from actual execution performance.

Trading / Educational Disclaimer

This article is for educational and research purposes only. Trading prediction markets involves market, liquidity, execution, model, and capital risk. No strategy discussed here guarantees profit.

Conclusion

The interesting property of Polymarket TWAP pricing is not simply that the reference is averaged. It is that averaging creates memory.

Top comments (0)