Skip to content

Incorporating Signals into Optimal Trading

Authors: Charles-Albert Lehalle, Eyal Neuman Institution: Capital Fund Management; CFM-Imperial College Institute; Imperial College London Year: 2017 (arXiv April 2017; published Finance and Stochastics 23(2), 2019) arXiv: 1704.00847 Categories: q-fin.TR, q-fin.MF

Plain-language abstract

Classical optimal-execution theory (Almgren-Chriss, Bertsimas-Lo, Gatheral-Schied-Slynko, Cartea-Jaimungal) models a trader who wants to minimise trading costs while offloading a large inventory. These frameworks typically assume the underlying price is a martingale — no predictive information. But real traders run short-horizon signals, and they rationally tilt their schedule when the signal favours or opposes their direction. This paper plugs a generic Markovian signal into the Gatheral-Schied-Slynko framework (transient market impact with fuel constraint), proves existence and uniqueness of the optimal strategy, and derives a closed-form solution for the Ornstein-Uhlenbeck signal + exponentially-decaying market impact case — the standard assumption for order-book-imbalance (OBI) signals. Empirically the paper calibrates OBI on NASDAQ OMX (13 stocks, 9 months, ~9B trades) and shows HFT market makers and proprietary traders already condition their trading rate on it. The OU model for OBI is empirically supported.

Key contributions

  1. Signal-in-GSS theorem (existence & uniqueness) — For a generic càdlàg Markov signal \(I\) and strictly positive-definite impact kernel \(G\), the optimal deterministic admissible execution strategy for the cost functional
\[\mathcal{C}(X) = \mathbb{E}\left[\int_0^T I_s X_s ds + \frac{1}{2}\int_0^T\!\int_0^T G(|t-s|) dX_t dX_s + \phi \int_0^T X_t^2 dt\right]\]

exists and is unique. Characterised as the solution to an integral equation (Theorems 2.3, 2.4 in the paper).

  1. Closed-form for OU signal + exponential impact (Corollary 2.7) — When \(I\) is Ornstein-Uhlenbeck with mean-reversion rate \(\gamma\) and volatility \(\sigma\), and \(G(t) = \kappa \rho e^{-\rho t}\) (Obizhaeva-Wang-style decay), the optimal schedule \(X^*_t\) is linear in both initial inventory \(x\) and initial signal value \(\iota\), plus a time-varying drift from OU dynamics. Reduces to Obizhaeva-Wang (2013) when \(\iota = 0\).

  2. Asymptotic equivalence of GSS → CJ frameworks — As the impact decay rate \(\rho \to \infty\), the transient impact kernel converges to Dirac delta (instantaneous impact), the singular GSS strategy's jumps vanish, and the optimal schedule becomes a smooth function matching the Cartea-Jaimungal framework. First bridge between the two literatures in a signal-enriched setting.

  3. Non-monotone optimal strategies — With a signal, the optimal execution is not necessarily monotone in inventory. A seller may temporarily buy when the signal points the wrong way. This opens the possibility of transaction-triggered price manipulation within a valid model. An open question: what conditions on \(G\) and \(I\) prevent this?

  4. Empirical validation of OBI as an OU signal — 9 months, 13 stocks, NASDAQ OMX. OBI is predictive of the next price move, exhibits mean-reverting behaviour consistent with an OU process, and calibration gives concrete \(\gamma\), \(\sigma\) estimates.

  5. Evidence practitioners use OBI — Using NASDAQ OMX's counterparty-identified trade data, the authors classify participants into four categories (global investment banks, institutional brokers, HFT market makers, HFT proprietary traders) and show that HFT participants systematically condition their trading rate on OBI, while long-only investors and brokers do not. Direct evidence that the signal-in-execution model is not just theoretical.

  6. Time-inconsistency discussion — Transient impact creates a control-theoretic inconsistency: the optimal strategy on \([0, T]\) computed at \(t=0\) is not in general the same as the concatenation of optimal strategies on \([t, T]\) computed at \(t\). Three practical options discussed: commit to the \(t=0\) optimal, re-plan continuously (approximate), or use the (CJ) instantaneous-impact limit (time-consistent).

Method summary

Model setup

Asset price decomposes as \(P_t = M_t + \int_0^t I_s ds\), where \(M\) is a martingale and \(I\) is the signal (drift component). Trader's visible price with transient impact:

\[S_t \;=\; P_t \;+\; \int_0^t G(t - s) \, dX_s\]

where \(X_t\) is inventory held at time \(t\) and \(G\) is the impact decay kernel. Fuel constraint: \(X_T = 0\) (liquidate fully).

Cost functional (GSS + signal)

\[\mathcal{C}(X) \;=\; \mathbb{E}\left[\int_0^T I_s X_s ds + \frac{1}{2}\int_0^T\!\int_0^T G(|t-s|)\, dX_t\, dX_s + \phi \int_0^T X_t^2 dt\right]\]

The three terms: - \(\int I_s X_s ds\) — the signal-inventory term. Having inventory \(X_s\) during a signal \(I_s > 0\) is profitable for a buyer (price going up) or costly for a seller. - \(\frac{1}{2} \iint G(|t-s|) dX_t dX_s\) — impact cost. Quadratic form in the flow \(dX\). - \(\phi \int X_t^2 dt\) — inventory risk penalty.

Optimal schedule for OU + exponential impact

With \(I_t\) OU and \(G(t) = \kappa\rho e^{-\rho t}\), the optimal strategy has the form:

\[X^*_t \;=\; b_0(t) \cdot x \;+\; b_1(t) \cdot I_0 \;+\; b_2(t) \cdot \int_0^t e^{-\gamma(t-s)} dW_s\]

for explicit functions \(b_0, b_1, b_2\) depending on \(\gamma, \sigma, \kappa, \rho, T\). Linear in initial inventory \(x\) and initial signal \(I_0\). Jumps occur at \(t = 0\) and \(t = T\) (singular part of the strategy), continuous in between.

Cartea-Jaimungal limit

In the \(\rho \to \infty\) (impact instantaneous) limit with instantaneous-impact kernel \(G(dt) = \kappa \delta_0\), the framework reduces to a standard HJB problem. With terminal penalty \(\varrho X_T^2\) instead of the fuel constraint:

\[r^*_t \;=\; \frac{v_1(t, I_t) - 2 v_2(t) X_t}{2\kappa}\]

where \(v_1, v_2\) solve a system of ODEs (Proposition 3.1). Matches Cartea-Jaimungal with a signal.

Empirical results

Data

  • NASDAQ OMX (Nordic European exchange), 9 months.
  • 13 stocks, > 9 billion transactions.
  • CFM proprietary order-book database merged with NASDAQ OMX's counterparty-identified trade tape.

OBI as a signal

Define OBI at the best quotes. Shown: - Positive correlation with next-move direction. - Mean-reverting; OU fit passes standard diagnostics. - Typical calibration range: \(\gamma \in [0.1, 1]\) per minute, \(\sigma\) stock-dependent.

Who uses OBI?

Average OBI value just before each trade, split by counterparty class:

Class Conditional on signal?
Global investment banks No / weak
Institutional brokers No
HFT market makers Yes — strongly
HFT proprietary traders Yes — strongly

Fig. 9 of the paper shows trading speed as a function of OBI within a 10-minute window — monotone-increasing, matching the theory's prediction that signal-aware strategies condition on current signal state.

Limitations

  • Deterministic strategies only. Theorems cover strategies using \(I_0\) alone, not signal-adaptive strategies (Remark 2.9). The latter is explicitly flagged as open.
  • Price manipulation possibility is not resolved. Non-monotone strategies can arise; conditions that rule out manipulation are an open problem.
  • Linear cost structure. Impact enters quadratically (linear marginal impact per unit of flow). Realistic impact is concave (square-root law) at larger sizes.
  • Specific kernel pair. Closed form is only for OU signal + exponential decay kernel. Other combinations (power-law decay, jump-diffusion signals) are not treated explicitly.
  • Time inconsistency. The framework is acknowledged time-inconsistent under transient impact; three workarounds discussed but no principled resolution.
  • European mid-cap universe. Empirical validation is 13 NASDAQ-OMX stocks; US large-caps or futures may behave differently.
  • Counterparty classification. The four-class taxonomy depends on NASDAQ-OMX metadata that is no longer published (post-2014).

Connections to other wiki pages

  • Framework: optimal-execution — extends with an explicit signal term in the cost functional.
  • Signal source: order-flow-imbalance — the OU model for OBI connects the execution theory back to the empirical OFI literature.
  • Impact primitive: propagator-model — the GSS transient-impact kernel is a propagator; the paper uses the exponential decay variant (Obizhaeva-Wang). The paper's empirical results on OBI mean-reversion parallel the propagator's \(\tau^{-\beta}\) decay story.
  • Method: signal-aware-optimal-execution (new) — the general framework this paper introduces.
  • Natural companion: cross-impact-ofi-equity-markets — OBI-driven execution naturally extends to multi-asset settings using cross-asset OFI as the signal.
  • Cartea-Jaimungal: referenced extensively but not (yet) in this wiki as a dedicated paper page. Their framework is the \(\rho \to \infty\) limit.

Signal-design takeaways

For someone building an L3 / LOB-snapshot signal specifically for execution:

  1. Model your signal as an OU process. OBI and related features are empirically mean-reverting; an OU calibration plugs directly into the closed-form schedule of Corollary 2.7.
  2. Estimate signal \(\gamma\) (mean-reversion rate) before anything else. The optimal tilt in the schedule is a function of \(\gamma\) and the fuel horizon \(T\) — mis-estimating \(\gamma\) rescales the effective signal size incorrectly.
  3. Use the signal even if you can only use its value at \(t=0\). The deterministic-strategy theorem guarantees the closed-form schedule with only initial-signal information is a valid optimiser. Adaptive updating is better but not strictly necessary.
  4. Size impact carefully. The exponential-decay rate \(\rho\) interacts with \(\gamma\) — if \(\rho \ll \gamma\), the signal changes faster than impact resolves, and the schedule flattens. If \(\rho \gg \gamma\), you're effectively in the CJ instantaneous-impact regime.
  5. Watch for non-monotone orders. If your closed-form says to buy while liquidating, that's mathematically optimal but may be blocked by pre-trade risk controls, cross-venue best-execution rules, or internal compliance. Bake a monotonicity constraint in if needed — at a cost.
  6. HFTs already do this. If you're a slower trader, the signal is already partially priced in by HFT flow that reacts faster than you can. Check the post-HFT residual information content before relying on OBI too heavily.