State Space Model

The core innovation is the separation of the true system state from what we can measure. The state vector $x(t)$ captures all relevant internal dynamics—position, velocity, market regime—while the observation $y(t)$ is a noisy, partial measurement. This latent structure allows the model to infer hidden drivers of observable behavior, making it ideal for filtering tasks like estimating true asset value from noisy price ticks.

Classical SSMs assume a linear transition matrix $A$ and observation matrix $C$ that remain constant over time. This LTI property enables powerful analytical tools:

• Impulse response analysis for shock propagation
• Frequency domain decomposition via transfer functions
• Closed-form solutions for forecasting While restrictive, LTI models underpin the Kalman filter and provide interpretable structure before introducing non-linear extensions.

The High-order Polynomial Projection Operator (HiPPO) provides a mathematically principled way to initialize the state transition matrix $A$. It compresses continuous signal history into a fixed-size state by projecting onto Legendre polynomials, ensuring the state optimally retains information over arbitrarily long timescales. This breakthrough, central to the S4 and Mamba architectures, solves the vanishing memory problem that plagued earlier recurrent models.

Continuous-time SSMs must be converted to discrete form for digital computation. The bilinear (Tustin) transform maps the continuous matrices $(A, B, C, D)$ to discrete equivalents while preserving stability properties. The step size $\Delta$ controls the temporal resolution:

• Smaller $\Delta$: finer granularity, higher computational cost
• Larger $\Delta$: coarser steps, faster processing This discretization step is now learned end-to-end in models like Mamba, allowing the system to adapt its effective sampling rate to the input.

Traditional SSMs use time-invariant dynamics, meaning the state transition is identical for every input token. Mamba introduces a selective scan where matrices $B$, $C$, and $\Delta$ become functions of the input:

• The model learns to focus on relevant tokens
• Irrelevant context is filtered out at the state level
• Achieves transformer-quality attention with linear complexity $O(N)$ This selectivity bridges the gap between efficient SSMs and the context-aware power of attention mechanisms.

The Kalman filter is the canonical application of SSMs to noisy observation streams. It performs recursive Bayesian updating in two steps:

1. Predict: Use the state transition $A$ to forecast the next state and its uncertainty
2. Update: Incorporate the new observation, weighted by the Kalman gain The gain dynamically balances trust between the model prediction and the measurement based on their relative uncertainties. This framework is foundational for real-time volatility estimation and order flow filtering in quantitative trading systems.

The Kalman filter—the canonical inference algorithm for linear-Gaussian state space models—is widely used to estimate time-varying hedge ratios in pairs trading. Unlike static regression, it treats the spread relationship as a latent state evolving via a random walk, adapting to structural breaks in correlation.

• Observation equation: Asset Y price = hedge ratio × Asset X price + noise
• State equation: Hedge ratio evolves as a random walk
• Outputs a dynamic beta that updates with each new tick, minimizing variance

State space models decompose asset returns into a latent time-varying volatility process and an observation noise term. The Heston model and its extensions treat variance as a mean-reverting latent state (CIR process), enabling options pricing that captures the volatility smile.

• Latent state: Instantaneous variance follows a square-root diffusion
• Observation: Log-returns with variance equal to the latent state
• Estimated via particle filters or quasi-maximum likelihood for non-Gaussian cases

Markov-switching state space models allow market parameters to shift between discrete latent regimes—bull, bear, and sideways—with stochastic transition probabilities. This captures volatility clustering and structural breaks that linear models miss.

• Discrete latent state: Market regime indicator
• Continuous latent state: Time-varying parameters within each regime
• Used for tactical asset allocation and risk budgeting that adapts to regime changes

When market dynamics are non-Gaussian—fat-tailed returns, jumps, or nonlinear price impact—particle filters (sequential Monte Carlo) approximate the latent state posterior using weighted samples. They track multi-modal distributions that Kalman filters cannot represent.

• Resampling step prevents weight degeneracy in volatile regimes
• Used for real-time bid-ask spread estimation and liquidity tracking
• Handles sudden market dislocations without Gaussian assumptions

Dynamic factor models cast asset returns as a linear combination of latent common factors plus idiosyncratic noise, all evolving through time. The state space framework estimates the factors and loadings jointly, adapting to rotating market themes.

• Observation: Cross-section of asset returns
• Latent state: Factor values and time-varying loadings
• Enables statistical arbitrage by trading residuals after stripping common factor exposure

The foundational recursive Bayesian estimator for linear dynamic systems. It operates in two steps: a predict step that propagates the state estimate forward using a transition model, and an update step that corrects the estimate using noisy observations. The Kalman gain optimally weights the prediction against the measurement based on their covariances.

• Assumes linear Gaussian dynamics
• Computes exact posterior distributions analytically
• Core to GPS navigation, tracking systems, and control theory
• The workhorse of state estimation for over 60 years

A discrete-state SSM where the latent variable takes one of N possible states and transitions follow a Markov chain. Observations are emitted according to state-dependent probability distributions. The Baum-Welch algorithm (a special case of EM) learns parameters, while the Viterbi algorithm finds the most likely state sequence.

• States are categorical, not continuous
• Widely used in speech recognition and bioinformatics
• Limited by the Markov assumption on latent states
• Predecessor to more expressive continuous-state models

A nonlinear SSM where the state transition is a learned neural network function. The hidden state h_t summarizes all past inputs and is updated recursively. While powerful in theory, vanilla RNNs suffer from the vanishing/exploding gradient problem, limiting their ability to capture long-range dependencies.

• LSTM and GRU variants add gating mechanisms
• Gating creates a form of selective state updating
• Historically dominant before the Transformer revolution
• Can be viewed as a discretized continuous-time SSM

A Monte Carlo method for state estimation in nonlinear, non-Gaussian SSMs. Instead of maintaining a parametric belief, it represents the posterior distribution as a set of weighted samples called particles. Sequential Importance Resampling (SIR) propagates and reweights particles at each time step.

• Handles arbitrary dynamics and observation models
• Computational cost scales with number of particles
• Used in robotics (SLAM) and target tracking
• Also known as a Sequential Monte Carlo method