Signal-Triggered Trading Protocol

• Signal-Triggered Trading Protocol is a framework where trade executions are governed by predefined statistical signals and control laws, integrating threshold events and continuous feedback.
• The protocol utilizes advanced methods such as reinforcement learning, smart contracts, and cryptographic proofs to validate signals and manage dark liquidity risks.
• It bridges discrete thresholding with continuous control and deep learning strategies, offering versatile models for optimal trade execution and risk management.

Searching arXiv for papers on signal-triggered trading protocols and closely related execution frameworks. {"query":"signal-triggered trading protocol optimal trading using signals dark liquidity fisher poisson deep reinforcement learning trading signal execution arXiv","max_results":10} Searching more targeted queries for core papers on signal-threshold execution, dark liquidity signaling detection, and RL-based signal execution. {"query":"(March et al., 2018, Zovko, 2017, Nagy et al., 2023, Lehalle et al., 2017, Futter et al., 2023, Kim et al., 4 Feb 2026) arXiv","max_results":10} A signal-triggered trading protocol is a trading or execution procedure in which the arrival, magnitude, sign, statistical significance, or verification status of a signal determines when orders are submitted, how inventory is adjusted, which venue is accessed, or whether execution is authorized at all. In the literature, the triggering object ranges from a thresholded scalar predictor, a per-fill statistical test for information leakage, and a Markovian or order-flow signal, to a GRU-filtered latent regime probability, a rough-path signature, a semantically filtered lead-lag relation, or a cryptographically verified message (March et al., 2018, Zovko, 2017, Lehalle et al., 2017, Nagy et al., 2023, Kim et al., 4 Feb 2026, Kocaoğullar et al., 2021).

1. Signal, trigger, and control variable

The basic architecture is consistent across otherwise dissimilar models. A protocol first defines a signal, then maps that signal into a trigger condition or control law, and finally links the trigger to an execution mechanism. In De March and Lehalle’s formulation, the trader observes a scalar signal $S_n$ at discrete wake-up times and executes one unit if and only if $S_n \ge \theta_n$, so the trigger is a threshold event and the control variable is the threshold $\theta_n$ itself (March et al., 2018). In the dark-liquidity setting of “How Fisher catches Poisson in the Dark,” the signal is not a forecast of price direction but a one-sided $p$-value attached to a dark fill, computed from the waiting time to the next lit print; the trigger is an alert when a combined $p$-value falls below $\alpha$ (Zovko, 2017). In the prediction-market protocol of “LLM as a Risk Manager,” the trigger is mechanical: if the leader’s relative move satisfies $|r_{L,t}|>\theta$, the follower is traded with fixed size and fixed holding horizon (Kim et al., 4 Feb 2026).

Other formulations replace hard thresholding by continuous signal-adaptive control. In “Incorporating Signals into Optimal Trading,” the unaffected mid-price satisfies $dP_t = Y_t\,dt + \sigma^P dW_t^P$, and the control is a trading rate $v_t$ chosen under transient impact (Lehalle et al., 2017). In “Optimal Trading with Signals and Stochastic Price Impact,” the execution rate is modulated both by a predictive signal $\mu_t$ and by a fast mean-reverting temporary-impact factor $S_n \ge \theta_n$0 (Fouque et al., 2021). In “Fast and Slow Optimal Trading with Exogenous Information,” the high-frequency trader’s optimal signal-adaptive rate is of feedback form $S_n \ge \theta_n$1 (Cont et al., 2022). In reinforcement-learning formulations, the signal enters the state and the trigger is implicit in the learned policy rather than a single explicit inequality (Nagy et al., 2023, Macrì et al., 31 Oct 2025).

Taken together, these works suggest that “signal-triggered” is not synonymous with “signal-thresholded.” It includes Bernoulli thresholding, continuous feedback control, sequential decision rules in event time, and authorization layers that permit execution only when statistical or cryptographic conditions are met.

2. Statistical triggering and real-time toxicity detection

A particularly explicit signal-triggered protocol appears in dark-pool execution. Lit-market trade arrivals are modeled as a Poisson point process with locally time-varying intensity $S_n \ge \theta_n$2, so waiting times $S_n \ge \theta_n$3 are exponential. After a dark fill at time $S_n \ge \theta_n$4, the protocol measures $S_n \ge \theta_n$5, the time to the next lit print, estimates the local rate from the $S_n \ge \theta_n$6 preceding lit waiting times, and assigns the fill a one-sided $S_n \ge \theta_n$7-value

$S_n \ge \theta_n$8

Low $S_n \ge \theta_n$9 means a surprisingly short wait and is treated as evidence that the dark fill may have triggered the lit print. Evidence across fills is then aggregated with Fisher’s method,

$\theta_n$0

with $\theta_n$1 under the null. If the combined $\theta_n$2-value falls below $\theta_n$3, the protocol triggers a signalling alert (Zovko, 2017).

The decision layer is deliberately operational. The paper describes four post-alert responses: immediately suspend dark participation on a venue-size bucket; impose or increase a minimum fill-size threshold; reduce participation rate; or apply a directional rule in which same-direction lit prints motivate scaling down while opposite-direction prints may justify slowing down or speeding up depending on strategy. Calibration is performed by backtesting over $\theta_n$4, by inspecting the empirical distribution of $\theta_n$5, and by sensitivity tests on false alarms versus missed signalling. Typical choices are $\theta_n$6, $\theta_n$7–$\theta_n$8, and $\theta_n$9 (Zovko, 2017).

The empirical motivation is the protocol’s central claim. In 100 highly-traded LSE names over Apr–Oct 2016, individual dark fills with a single $p$0 suffered approximately $p$1 higher 5-second post-fill slippage than fills with $p$2. Grouping $p$3 fills and using $p$4, executions flagged as signalling within their first 5 fills exhibited approximately $p$5 worse arrival-price slippage than non-signalling ones. By contrast, a classical slippage-tracking $p$6-test would require $p$7 fills to reach statistical significance (Zovko, 2017).

A second statistical protocol appears in prediction markets, where the trigger is based on leader–follower structure rather than venue toxicity. Prices $p$8 are converted to log-odds, pairwise Granger-causality tests are run with lags $p$9, the top $p$0 directed pairs are retained, and an LLM semantic filter re-ranks them to select the top $p$1. Trading then follows a fixed rule: if $p$2 with default $p$3, enter the follower at the next day’s price, size the position at $p$4, and hold for $p$5 trading days. Across rolling evaluations on Kalshi Economics markets, the hybrid approach increased win rate from $p$6 to $p$7 and reduced the average magnitude of losing trades from 649 USD to 347 USD (Kim et al., 4 Feb 2026). Here the signal-triggered component is entirely mechanical, while the semantic layer functions as a pre-trade risk filter.

3. Threshold control and continuous-time optimal execution

De March and Lehalle formulate the canonical threshold protocol in discrete time. At each observation time $p$8, the trader fixes $p$9 and buys one unit iff $\alpha$0, so

$\alpha$1

Under i.i.d. assumptions on $\alpha$2, a continuous-time rescaling yields

$\alpha$3

and the Inventory Asymptotic Behaviour theorem implies that $\alpha$4 is approximately Gaussian around the deterministic trajectory $\alpha$5 with variance of order $\alpha$6. The associated control problem maximizes expected cumulative gain subject to a terminal inventory target, leading to an HJB equation in which the control is the execution probability $\alpha$7. In the deterministic limit $\alpha$8, the optimizer is

$\alpha$9

This formulation is notable because it treats the uncertainty induced by thresholding itself as the object of analysis and also provides an ex-post mapping from any speed-driven algorithm $|r_{L,t}|>\theta$0 to a threshold rule via $|r_{L,t}|>\theta$1 (March et al., 2018).

Lehalle and Neuman incorporate a Markovian signal into a transient-impact execution problem. The unaffected price satisfies $|r_{L,t}|>\theta$2, the visible mid-price is

$|r_{L,t}|>\theta$3

and inventory obeys $|r_{L,t}|>\theta$4. With continuous, bounded, strictly positive-definite impact kernel $|r_{L,t}|>\theta$5, there exists at most one deterministic minimizer, characterized by a Fredholm integral equation. In the special case $|r_{L,t}|>\theta$6, $|r_{L,t}|>\theta$7, and Ornstein–Uhlenbeck signal $|r_{L,t}|>\theta$8, the optimal liquidation path is singular: it has a Dirac jump at $|r_{L,t}|>\theta$9, a smooth component $dP_t = Y_t\,dt + \sigma^P dW_t^P$0, and a Dirac jump at $dP_t = Y_t\,dt + \sigma^P dW_t^P$1. As $dP_t = Y_t\,dt + \sigma^P dW_t^P$2, the strategy converges to a continuous Cartea–Jaimungal limit (Lehalle et al., 2017).

Fouque, Jaimungal, and Saporito add stochastic temporary impact. The execution rate admits the approximation

$dP_t = Y_t\,dt + \sigma^P dW_t^P$3

where

$dP_t = Y_t\,dt + \sigma^P dW_t^P$4

The signal $dP_t = Y_t\,dt + \sigma^P dW_t^P$5 shifts the base trading speed, while the fast mean-reverting friction factor $dP_t = Y_t\,dt + \sigma^P dW_t^P$6 rescales it multiplicatively (Fouque et al., 2021).

In “Optimal execution and speculation with trade signals,” the signal is a Meyer-$dP_t = Y_t\,dt + \sigma^P dW_t^P$7-field object $dP_t = Y_t\,dt + \sigma^P dW_t^P$8 that becomes observable just before the market-impact and liquidity shock materialize. This permits pre-emptive signal-based trades $dP_t = Y_t\,dt + \sigma^P dW_t^P$9, followed by post-shock state-based trades $v_t$0, within a Marcus-type SDE system and an HJB quasi-variational inequality (Bank et al., 2023). In the Stackelberg game of “Fast and Slow Optimal Trading with Exogenous Information,” the signal-adaptive high-frequency trader and the deterministic institutional investor are solved jointly; the HFT’s rule is linear feedback, while the major agent’s strategy is obtained from the resolvent of a Fredholm integral equation (Cont et al., 2022). These models show that a signal-triggered protocol can be an equilibrium object rather than a single-agent heuristic.

4. Limit-order-book learning, latent information, and path dependence

A discrete execution protocol of a different kind is developed in the ABIDES-based limit-order-book environment of “Asynchronous Deep Double Duelling Q-Learning for Trading-Signal Execution in Limit Order Book Markets.” The state includes time remaining, cash balance, inventory, a directional forecast $v_t$1, and a short history of best bid/ask prices, depths, and resting volumes. The action space has $v_t$2: $v_t$3 with $v_t$4 and $v_t$5. Reward combines log-return P&L and a directional term encouraging inventory alignment with the forecast, with a decaying weight so that training begins signal-following and ends P&L-focused. The architecture uses Deep Duelling Double Q-learning with the APEX asynchronous architecture, $v_t$6 CPU actors, $v_t$7 GPU learner, replay capacity $v_t$8, batch size $v_t$9, $\mu_t$0-step returns with $\mu_t$1, and target-network updates every $\mu_t$2 learner steps. The paper reports that the RL agent learns an effective trading strategy for inventory management and order placing that outperforms a heuristic benchmark having access to the same signal (Nagy et al., 2023).

A partial-information variant is given in “Deep reinforcement learning for optimal trading with partial information,” where the signal follows an Ornstein–Uhlenbeck process with regime-switching mean reversion, speed, and volatility. Three DDPG-based algorithms are compared: hid-DDPG, prob-DDPG, and reg-DDPG. The key distinction is informational structure. hid-DDPG feeds GRU hidden states directly into the actor-critic system; reg-DDPG uses a GRU-based forecast $\mu_t$3; prob-DDPG uses GRU-estimated posterior regime probabilities $\mu_t$4. In synthetic tests, prob-DDPG dominates: with $\mu_t$5 switching only, average cumulative rewards after $\mu_t$6 steps and $\mu_t$7 test episodes are $\mu_t$8 for prob-DDPG, $\mu_t$9 for hid-DDPG, and $S_n \ge \theta_n$00 for reg-DDPG; with $S_n \ge \theta_n$01 all switching, the values are $S_n \ge \theta_n$02, $S_n \ge \theta_n$03, and $S_n \ge \theta_n$04, respectively. In a one-day empirical pair-trading test on SMH versus INTC, prob-DDPG yields $S_n \ge \theta_n$05, while hid-DDPG and a Z-score benchmark give $S_n \ge \theta_n$06 and $S_n \ge \theta_n$07 (Macrì et al., 31 Oct 2025).

Futter, Horvath, and Wiese’s “Signature Trading” pushes the notion of signal-triggering beyond Markov state variables. The tradable price process $S_n \ge \theta_n$08 and exogenous signal $S_n \ge \theta_n$09 are time-augmented into $S_n \ge \theta_n$10, lifted to the truncated signature $S_n \ge \theta_n$11, and mapped to positions by linear functionals: $S_n \ge \theta_n$12 Expected return and variance become $S_n \ge \theta_n$13 and $S_n \ge \theta_n$14, so the mean-variance problem has explicit solution

$S_n \ge \theta_n$15

with a variance constraint generating a “Signature Efficient Frontier.” Momentum and pair-trading examples are given, and higher-order signature terms encode path dependence without abandoning closed-form quadratic optimization (Futter et al., 2023).

5. Cryptographic, smart-contract, agentic, and communication-theoretic protocols

Signal-triggering can also be implemented as protocol logic rather than purely as a quantitative control law. In “A Smart-Contract to Resolve Multiple Equilibrium in Intermediated Trade,” a smart contract receives messages and algorithmically sends trading instructions in a broker-dealer market with multiple equilibrium. The protocol implements broker-dealer joint profit maximization as a Nash equilibrium and relies on agent commitments to follow the protocol, selective privacy of information, structured timing of trade offers and acceptances, and trust that the smart contract executes the correct algorithm. Commitment is achieved by a legal contract or contingent deposit, privacy by fully homomorphic encryption, and verifiability by appending the smart contract to a public blockchain (Aronoff et al., 28 May 2025).

ChainBot provides a concrete on-chain architecture for signal-triggered algorithmic trading. Public price data are fetched from an oracle, an off-chain bot computes a private Bollinger-Band signal, a zero-knowledge proof certifies that the signal was computed correctly, and only then does the on-chain contract authorize trade relay to a privacy-preserving layer-2 DEX. The design partitions the strategy into on-chain balances and verification, off-chain signal computation and proof generation, and layer-2 private execution. The paper reports returns up to $S_n \ge \theta_n$16 the buy-and-hold baseline, an end-to-end average execution time of $S_n \ge \theta_n$17 seconds across $S_n \ge \theta_n$18 runs, and the observation that trading frequency does not significantly affect rate of return and Sharpe ratio (Kocaoğullar et al., 2021).

AgenticAITA replaces “signal then execute” by a deliberative pipeline. Its Adaptive Z-Score Trigger Engine defines $S_n \ge \theta_n$19, computes

$S_n \ge \theta_n$20

with $S_n \ge \theta_n$21, and triggers when $S_n \ge \theta_n$22 or $S_n \ge \theta_n$23. Triggered events enter a mutex-based Inference Gating Protocol with cooldown $S_n \ge \theta_n$24 seconds, after which an Analyst agent, Risk Manager agent, and Executor agent interact through typed JSON contracts. The Risk Manager applies four hard gates: signal must be “long” or “short,” confidence must be at least $S_n \ge \theta_n$25, $S_n \ge \theta_n$26, and size must not exceed 500 USD. A Correlation-Break Diversification score

$S_n \ge \theta_n$27

is injected into the Analyst prompt. In a five-day autonomous dry run, the system recorded $S_n \ge \theta_n$28 zero-intervention invocations across $S_n \ge \theta_n$29 assets, $S_n \ge \theta_n$30 Analyst self-abstentions, $S_n \ge \theta_n$31 hard-gate rejections, agentic friction $S_n \ge \theta_n$32, $S_n \ge \theta_n$33 executed dry-run trades, benchmark alpha versus BTC buy-and-hold of $S_n \ge \theta_n$34 percentage points, and win rate $S_n \ge \theta_n$35 with binomial $S_n \ge \theta_n$36 (Letteri, 1 May 2026).

At the communication layer, “Communication Strategies for Low-Latency Trading” studies a binary signal transmitted over a noisy channel. “Buy” and “Sell” are encoded by repetition codes of length $S_n \ge \theta_n$37 over a binary symmetric channel with crossover probability $S_n \ge \theta_n$38; majority decoding yields block-error probability

$S_n \ge \theta_n$39

Here $S_n \ge \theta_n$40 is itself the strategic variable: shorter codes reduce latency but increase decoding error. The resulting game has either a unique symmetric tie equilibrium or two asymmetric Nash equilibria with different winners, depending on the noise regime (Karzand et al., 2015).

6. Misconceptions, strategic tensions, and open directions

A common simplification is to treat signal-triggered trading as a rule of the form “trade when signal exceeds threshold.” The literature is broader. There are hard-threshold rules based on $S_n \ge \theta_n$41, $S_n \ge \theta_n$42, or $S_n \ge \theta_n$43, but there are also continuous feedback laws, HJBQVI-based impulse controls, learned LOB policies, and linear functionals on path signatures (March et al., 2018, Kim et al., 4 Feb 2026, Letteri, 1 May 2026, Bank et al., 2023, Nagy et al., 2023, Futter et al., 2023).

Another misconception is that post-trade slippage is an adequate real-time proxy for information leakage. The dark-liquidity literature explicitly disputes this, arguing that slippage estimation needs a large number of fills and misses causality between dark fills and lit prints; the Fisher-Poisson protocol is intended precisely to detect signalling on a per-fill basis rather than after hundreds of observations (Zovko, 2017).

Strategic tension is also intrinsic rather than incidental. In the fast-slow Stackelberg model, the high-frequency trader may behave cooperatively or predatorily depending on the tradeoff between order-flow and the trading signal (Cont et al., 2022). In low-latency communication, identical channels can still generate multiple Nash equilibria (Karzand et al., 2015). In intermediated trade, a smart contract is proposed specifically to resolve multiple equilibrium by fixing a verifiable timing and instruction mechanism (Aronoff et al., 28 May 2025). These results place signal-triggered trading within market design and game theory, not only within statistical prediction.

Evaluation remains highly context-specific. Reported gains come from heterogeneous environments: 100 highly-traded LSE names over Apr–Oct 2016, Kalshi Economics markets over Oct 2021–Nov 2025, replayed NASDAQ limit-order-book data in ABIDES, SMH–INTC pair trading at 1-second frequency, and a five-day autonomous crypto dry run (Zovko, 2017, Kim et al., 4 Feb 2026, Nagy et al., 2023, Macrì et al., 31 Oct 2025, Letteri, 1 May 2026). This suggests that a fully general theory would have to integrate signal quality, market impact, equilibrium feedback, privacy, and execution infrastructure in a single framework. The existing literature instead provides a set of rigorously specified but domain-dependent templates for linking a signal to a trading action.