Adaptive Execution Under Volatility Spikes

When to Use

Use this overlay before each child-order decision in a TWAP, VWAP, or POV schedule when a trusted real-time feed supplies a short-horizon volatility signal. It classifies the signal into NORMAL, HIGH_VOLATILITY, or CRITICAL_SHOCK and returns bounded execution parameters.

The engine is deliberately advisory. It does not place, cancel, replace, or route orders and it does not enforce broker, exchange, or portfolio risk controls.

When NOT to Use

• Do not use it as a substitute for broker or exchange pre-trade controls, trading-pause handling, limit-up/limit-down validation, or a portfolio kill switch.
• Do not use it when the volatility feed is stale, unavailable, improperly calibrated, or mixes incompatible instruments or sessions.
• Do not use it as a standalone market-impact, liquidity, best-execution, or smart-order-routing model; volatility alone does not measure spread, depth, toxicity, or venue availability.
• Do not route an order solely because this engine returns NORMAL; the parent scheduler and independent risk gates must still approve it.

Prerequisites

• Python 3.10+.
• A real-time, instrument-specific volatility signal with documented units and timestamp/freshness checks performed by the caller.
• A parent execution scheduler and an EMS capable of idempotent cancel/replace requests.
• Independent pre-trade controls for quantity, notional, price collars, credit, position, and venue trading status.
• A tested operational procedure for halting, cancelling working orders, alerting, and manually or explicitly resuming a parent order.

Workflow

1. Validate configuration: Construct AdaptiveVolatilityConfig with participation in [0, 1], a positive base child size, non-negative offsets, finite thresholds, and critical > high. Leave enabled=True: enabled=False is a bypass that returns the base parameters and skips volatility validation entirely, not a safe default. The engine re-validates the config on every call, so a live edit or partial reload fails closed rather than routing on out-of-range bounds.
2. Validate the signal upstream: Confirm instrument identity, session, timestamp freshness, and calculation method. Pass a finite numeric current_volatility value to evaluate.
3. Evaluate before routing: Call AdaptiveExecutionUnderVolatilitySpikesEngine.evaluate(market_data) for every child-order decision, using one engine instance per instrument and parent order. Treat MarketDataValidationError or other validation failures as a safety event, not as normal-market input; a failed evaluation leaves engine.current_regime at UNKNOWN rather than at the regime of the last successful call, and a newly constructed engine reports UNKNOWN until an evaluation succeeds.
4. Apply the decision: branch on halt_trading before reading any numeric field.

• NORMAL: use configured participation, child size, and normal offset.
• HIGH_VOLATILITY: use half the configured participation and child size, with the high-volatility offset. Re-check all EMS and venue price/quantity limits.
• CRITICAL_SHOCK: do not submit new orders. Cancel all working orders for the parent using stable client/order identifiers, record the cancellation outcome, and alert operations.

limit_offset_bps is a distance away from the aggressive side of your reference price — a buy limit at ref * (1 - bps / 10_000), a sell limit at ref * (1 + bps / 10_000). The larger high-volatility offset is therefore the more passive one. Map the sign explicitly if your EMS defines offsets as aggressiveness.

1. Recover explicitly: Keep the parent paused after a critical shock until the external recovery policy is satisfied. Revalidate feed freshness, venue status, risk limits, and outstanding-order state before resuming; do not infer recovery from one normal observation.
2. Observe and reconcile: Emit regime, input timestamp, decision timestamp, parent/order identifier, cancel results, exceptions, and parameter values to an auditable event stream. Reconcile the EMS state before every resume.

Common Pitfalls

• Missing-data defaulting: Treating absent volatility as 0.0 silently enables normal trading. The implementation raises MarketDataValidationError instead.
• Control substitution: Treating a strategy threshold as an exchange collar, trading halt, or regulatory control. Those controls remain outside this engine.
• Non-idempotent cancellation: Sending repeated cancel requests without stable identifiers or reconciling acknowledgements can leave the parent partially active.
• Stale or cross-instrument data: A numerically valid signal can still be unsafe when its timestamp, instrument, session, or units are wrong.
• Threshold chatter: A single observation below a threshold is not a sufficient recovery policy. Use an external cooldown, hysteresis, or manual release procedure.
• Inverted offset convention: Applying limit_offset_bps as aggressiveness rather than passivity makes the high-volatility branch chase a dislocating book — the exact opposite of the intended protection. The halt decision zeroes every numeric field, so a caller that ignores halt_trading and prices off those zeros gets the most aggressive offset available.
• Bypass mistaken for a default: enabled=False disables the volatility validation along with the overlay. A NORMAL result from a disabled engine says nothing about the market; it was never measured.
• Shared engine instances: current_regime describes that instance's last evaluation only. One engine shared across symbols or parents reports whichever evaluated most recently, and concurrent calls on one instance race. Read ExecutionParameters.regime from the returned object for a per-decision value.
• Illusory liquidity: Smaller child orders and wider offsets do not guarantee fills or prevent slippage when displayed depth disappears.

Verification

Run the focused tests from the skill directory:

python -m unittest discover -s skills/adaptive-execution-under-volatility-spikes/scripts

The tests cover normal, boundary, negative-signal, high-volatility, child-size-floor, critical-shock halt-sentinel, disabled-bypass, missing-input, invalid-input, mapping-type, constructor-type, runtime config-mutation, and post-fault UNKNOWN state behavior. Before deployment, replay calibrated historical and synthetic shock scenarios and verify that the EMS cancels and reconciles working orders idempotently.

Related Skills

• execution-algo-twap-vwap-slicing
• execution-algorithm-kill-switch-integration
• order-placement-idempotency
• broker-api-idempotent-cancel-requests

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: HimanshuJ16
• Source: HimanshuJ16/Algo-Trading-Skills
• License: Apache-2.0
• Homepage: https://skills.himanshujangir.com

Install and usage instructions live in the source repository linked above.