strategy

mantis-strategy

Venue-agnostic strategy runtime primitives for the Mantis low-latency financial SDK.

Overview

This crate provides the building blocks for implementing trading strategies. All types are no_std compatible, use fixed-size arrays, and allocate nothing on the heap.

Prediction-market-specific types (YES/NO positions, settlement PnL, merge operations) belong in mantis-prediction, not here.

Components

Type	Purpose
Strategy	Event-driven trait — on_event(&mut self, event, intents) -> usize
OrderIntent	Post/Cancel/Amend intent, repr(C), Copy, flows through SPSC rings
Position	Signed inventory (SignedLots), VWAP entry, realized/unrealized PnL
OrderTracker	Fixed-size [Option<TrackedOrder>; 64] order state machine
QueueEstimator	L2 probabilistic queue position model (PowerProbQueueFunc)
ExposureView	Composes position + open orders for worst-case risk calculation
RiskLimits	Per-strategy risk configuration (position, capital, rate limits)
StrategyContext	Optional helper bundling engine + queue + orders + risk + positions

Strategy Trait

use mantis_strategy::{Strategy, OrderIntent, MAX_INTENTS_PER_TICK};
use mantis_events::HotEvent;

struct MyStrategy { /* ... */ }

impl Strategy for MyStrategy {
    const STRATEGY_ID: u8 = 0;
    const NAME: &'static str = "my-strategy";

    fn on_event(
        &mut self,
        event: &HotEvent,
        intents: &mut [OrderIntent; MAX_INTENTS_PER_TICK],
    ) -> usize {
        // Process event, return number of intents written
        0
    }
}

Design Decisions

• No generics on trait — OrderBook type and MAX instruments are implementation details inside each concrete strategy, not on the trait surface
• Associated consts — STRATEGY_ID and NAME are compile-time, not runtime methods
• Enum dispatch — bot binaries use enum dispatch (not dyn) for zero vtable overhead
• Lots vs SignedLots — unsigned for order sizes, signed for position inventory
• Fixed arrays — [Option<T>; 64] everywhere, no HashMap, no Vec, no heap
• Replay-friendly — strategy is a pure event→intent state machine. Same tape = same intents

Queue Position Model

The QueueEstimator uses a probabilistic model based on the PowerProbQueueFunc from hftbacktest:

• Cancels are biased toward the back of the queue (parameter n=2-3)
• Fill probability estimated via Poisson model: P[Poisson(take_rate * time) >= ahead_qty + order_qty]
• Take rate tracked per instrument per side via EWMA
• Calibrate n from live fill data

License

Apache-2.0