Backtests here replay venue data from Kalshi and Polymarket into NautilusTrader. The main things that move realized backtest performance beyond the raw venue data are:

• exchange fee models
• taker slippage assumptions
• engine behavior such as IOC handling, price rounding, cash-account limits, and AccountBalanceNegative

Fees

• Kalshi uses a nonlinear expected-earnings fee model
• Polymarket uses the current taker fee curve from venue metadata, plus CLOB fee-rate enrichment when the market payload itself still reports zero fees
• Polymarket maker fees are treated as zero; taker fees vary by market category and are not hardcoded to the older sports-vs-crypto heuristic
• if a venue reports zero fees for a market, the backtest also applies zero fees

Slippage

• shared prediction-market backtests default to a custom taker fill model
• non-limit orders get a deterministic one-tick adverse fill
• Polymarket uses the market's own tick size in trade-tick mode
• Kalshi uses one cent as the effective order tick for taker slippage
• limit orders keep default Nautilus matching behavior
• PMXT-backed Polymarket L2 backtests do not use the synthetic one-tick taker fill model; they replay historical OrderBookDeltas with book_type=L2_MBP and liquidity_consumption=True

Passive Orders And Queue Position

• public runners can now opt into Nautilus queue_position=True
• the public PMXT quote-tick runners in this repo now enable that heuristic by default, and the public trade-tick late-favorite passive runner now does too
• this is still a heuristic, not true venue queue reconstruction
• Kalshi and Polymarket trade-tick replay can use trade prints to move queue ahead estimates on passive limit orders
• PMXT quote-tick replay can also enable queue tracking, but that path replays L2 book updates and quotes rather than historical trade ticks, so fills still depend more heavily on book-level clears and price moves than on true trade consumption
• public MBP data does not expose hidden liquidity, exact priority inside a level, or venue-specific matching quirks

Latency

• public runners can now attach a static Nautilus latency model through the runner config
• the public PMXT quote-tick runners in this repo now ship with a static latency model enabled by default
• the current repo-layer surface is a static millisecond model with separate base, insert, update, and cancel delays
• this helps test whether a market-making or quote-chasing strategy only works because orders are assumed to land instantly

Limits

• repo-owned backtests keep cash-account risk checks enabled by default
• result payloads now distinguish the requested replay window from the loaded data window via planned_start, planned_end, loaded_start, loaded_end, coverage_ratio, and requested_coverage_ratio
• Kalshi public backtests here are trade-tick replay only
• Polymarket PMXT-backed backtests are full L2 order-book replay
• taker-heavy strategies that harvest tiny price changes can look much worse once fees and one-tick slippage are turned on
• PMXT L2 helps with taker modeling, but robust maker realism still needs L3 or MBO-style data

Vendor L2 Behavior

PMXT

• the loader prefers local filtered cache first, then raw sources in the order configured by the runner with local:, archive:, and relay:
• for the public PMXT runners in this repo, that usually means local raw mirror first, then the configured remote archive, then a raw mirror fallback
• the current shared relay direction is mirror-only, so the durable shared server path is raw parquet serving rather than server-side filtered-hour processing
• local PMXT filtered cache is enabled by default and grows with the number of unique (condition_id, token_id, hour) tuples you replay
• BACKTEST_ENABLE_TIMING=0 is the opt-out if you want a quieter PMXT run

For concrete timings and source tiers, see Vendor Fetch Sources And Timing.