Hawkes Process

The defining feature of a Hawkes process is its conditional intensity function, which spikes immediately after an event and then decays back toward a baseline level. This creates temporal clustering—a burst of activity makes further bursts more likely.

• The intensity is modeled as: λ(t) = μ + Σ g(t - tᵢ)
• μ (mu) is the baseline intensity (exogenous events)
• g(t) is the excitation kernel, typically an exponential decay: αe^(-βt)
• α (alpha) controls the jump size; β (beta) controls the decay speed
• The branching ratio n = α/β determines whether the process is subcritical (n < 1), critical (n = 1), or supercritical (n > 1)

Multivariate Hawkes processes model cross-excitation between different event types, capturing how one type of event triggers another. In limit order books, a large market buy order may excite subsequent sell limit orders and cancelations.

• Each event type k has its own baseline intensity μₖ
• The excitation matrix A encodes how events of type j excite type k
• Used to model order flow contagion across correlated assets
• Enables causal inference about lead-lag relationships in high-frequency data
• Captures the reflexivity between price changes and order flow imbalances

Hawkes process parameters are typically estimated using maximum likelihood estimation (MLE). The log-likelihood has a computationally tractable recursive form for exponential kernels, enabling efficient fitting to tick-level data.

• The log-likelihood decomposes into a compensator term and an event term
• Exponential kernels yield O(N) recursive computation
• Power-law kernels better capture long-memory in financial data but require numerical integration
• Regularization (L1/L2) prevents overfitting in high-dimensional multivariate models
• Goodness-of-fit assessed via residual analysis and the time-change theorem

Hawkes processes are a workhorse model for high-frequency financial data, capturing the clustering of trades, quote updates, and order book events that violate the Poisson assumption of independent arrivals.

• Trade arrival modeling: Captures bursts of trading activity during news events
• Order book dynamics: Models the interplay between limit orders, market orders, and cancelations
• Price jump prediction: Self-excitation signals impending volatility clusters
• Optimal execution: Informs strategies to minimize market impact during clustered activity
• Market making: Helps quote placement by anticipating short-term order flow intensity

The Hawkes process has a natural interpretation as a branching process, where each event is either an immigrant (arriving exogenously at rate μ) or an offspring (triggered by a previous event). This decomposition enables causal attribution of event clusters.

• The branching ratio n is the expected number of offspring per event
• Events can be probabilistically assigned to parent events via stochastic declustering
• Enables identification of exogenous shocks vs. endogenous cascades
• Critical for distinguishing informed trading from herding behavior
• Connects to Econophysics models of market reflexivity and feedback loops

Unlike a homogeneous Poisson process where events arrive independently at a constant rate, the Hawkes process introduces path-dependent excitation. This distinction is critical for accurately modeling clustered financial phenomena.

• Poisson: Constant intensity, independent increments, no clustering
• Hawkes: Time-varying intensity, dependent increments, generates clustering
• The Hawkes process reduces to a Poisson when α → 0 (no excitation)
• Over-dispersion: Hawkes processes produce variance greater than the mean
• Better captures volatility clustering and the empirical autocorrelation of absolute returns

The defining mechanism of a Hawkes process where each event arrival increases the conditional intensity function by a fixed amount, making subsequent events more likely in the near future. This creates temporal clustering without requiring external shocks.

• The excitation decays exponentially over time, governed by a kernel function
• Captures the empirical observation that trades and order book events occur in bursts
• Contrasts with a homogeneous Poisson process, where events are independent

The instantaneous rate of event arrival at time t, conditional on the entire history of past events. It is the central object of inference in a Hawkes process.

• Defined as: λ(t) = μ + Σ φ(t - tᵢ)
• μ (mu): The baseline intensity, representing exogenous background activity
• φ (phi): The excitation kernel, quantifying how past events at times tᵢ influence the current rate
• Directly models the time-varying nature of market activity

A critical stability parameter, often denoted by n, representing the expected number of direct descendant events triggered by a single parent event. It is the integral of the excitation kernel.

• n < 1 (Sub-critical): The process is stationary; event clusters eventually die out
• n = 1 (Critical): The process is at the boundary of stability
• n > 1 (Super-critical): The process is explosive; event rate grows without bound
• In finance, the branching ratio quantifies the endogeneity of market activity—the proportion of trades caused by other trades

An extension modeling multiple interacting event types, where events in one dimension can excite activity in another. Essential for capturing cross-asset and cross-order-type dynamics.

• Uses an excitation matrix to quantify mutual influence between dimensions
• Example: A large buy market order (Dimension 1) may excite subsequent sell limit orders (Dimension 2)
• Enables the modeling of lead-lag relationships and information flow between assets
• Parameter estimation typically uses maximum likelihood estimation (MLE) or moment-based methods

The parametric form of the excitation decay φ(t - tᵢ). The choice of kernel dictates the temporal shape of the self-exciting influence.

• Exponential Kernel: φ(s) = αe^(-βs). The most common choice due to its Markovian property, enabling fast simulation and estimation
• Power-Law Kernel: φ(s) = K/(s + c)^(1+θ). Captures long-memory effects observed in financial order flow, decaying more slowly than an exponential
• Sum of Exponentials: Approximates complex decay patterns with multiple timescales

Validating a fitted Hawkes model by checking whether its residuals behave like a unit-rate Poisson process. This is done via the random time change theorem.

• Transform event times tᵢ using the integrated conditional intensity: τᵢ = ∫₀ᵗⁱ λ(s) ds
• If the model is correct, the transformed times {τᵢ} should be a homogeneous Poisson process with rate 1
• Use a Q-Q plot against an exponential distribution or a Kolmogorov-Smirnov test to assess fit
• Essential for ensuring the model accurately captures the clustering structure before using it for prediction