12.3 Applications of state-space models and Kalman filtering

State-space models and Kalman filtering provide a framework for estimating hidden (unobserved) variables in systems that change over time. They combine a model of how the system evolves with noisy observations to produce optimal estimates of the true underlying state. This makes them valuable across finance, engineering, and the natural sciences, especially when data is multivariate, noisy, or contains gaps.

Compared to traditional approaches like ARIMA or exponential smoothing, state-space models handle more complex structures naturally: multiple related variables, time-varying parameters, and missing observations are all built into the framework rather than bolted on.

Applications and Evaluation of State-Space Models and Kalman Filtering

Applications of state-space models

Finance. State-space models are widely used in finance because many quantities of interest (true volatility, risk exposures, fair value) aren't directly observable. You observe market prices and returns, then use the model to infer what's happening underneath.

• Portfolio optimization — estimating how optimal asset allocations shift over time as market conditions change
• Asset pricing — estimating the fair value of stocks or bonds based on latent economic factors that drive returns
• Risk management — tracking evolving financial risks (market risk, credit risk) to inform hedging and capital decisions
• Volatility forecasting — predicting future price fluctuations for options pricing and trading strategies, where volatility itself is an unobserved state

Engineering. In engineering, the "state" is often a physical quantity like position, velocity, or orientation that you can only measure indirectly through noisy sensors.

• Control systems — representing and optimizing dynamic system behavior in robotics and aerospace
• Navigation and guidance — estimating position, velocity, and orientation of aircraft or spacecraft from sensor readings (GPS, accelerometers, gyroscopes) and fusing those measurements optimally
• Robotics — modeling how robots move and interact with their environment, enabling autonomous vehicles and industrial automation
• Signal processing — filtering and estimating signals buried in noise, with applications in speech recognition and image processing

Natural sciences. Many natural systems have hidden states (true atmospheric conditions, population sizes, disease prevalence) that are only partially observed.

• Weather forecasting — predicting atmospheric conditions by combining physical models of the atmosphere with sparse, noisy observations from weather stations and satellites
• Oceanography — modeling ocean currents, waves, and their interactions with climate systems
• Ecology — tracking population dynamics and species interactions (e.g., predator-prey relationships) where direct counts are unreliable
• Epidemiology — estimating the true spread of infectious diseases like COVID-19 or influenza from incomplete case reports, and evaluating intervention effectiveness

Forecasting with Kalman filtering

The Kalman filter operates on two equations that together define the state-space model:

• State equation: $x_t = F_t x_{t-1} + v_t$ — describes how the hidden state evolves over time. $F_t$ is the state transition matrix, and $v_t$ is process noise (random shocks to the system).
• Observation equation: $y_t = H_t x_t + w_t$ — links what you actually observe ($y_t$) to the hidden state. $H_t$ is the observation matrix, and $w_t$ is measurement noise.

The filter itself alternates between two steps at each time point:

Step 1: Prediction. Use the model to project forward from the last known state.

1. Predict the state: $\hat{x}_{t|t-1} = F_t \hat{x}_{t-1|t-1}$ This takes your best estimate from the previous time step and applies the transition model.
2. Predict the covariance: $P_{t|t-1} = F_t P_{t-1|t-1} F_t^T + Q_t$ This tracks how uncertain you are about the prediction. $Q_t$ is the process noise covariance, which adds uncertainty because the system is subject to random shocks.

Step 2: Update. Incorporate the new observation to correct the prediction.

1. Compute the Kalman gain: $K_t = P_{t|t-1} H_t^T (H_t P_{t|t-1} H_t^T + R_t)^{-1}$ This determines how much weight to give the new observation versus the prediction. When measurement noise ($R_t$) is large relative to prediction uncertainty, $K_t$ is small and you trust the prediction more. When $R_t$ is small, you trust the observation more.

2. Update the state: $\hat{x}_{t|t} = \hat{x}_{t|t-1} + K_t (y_t - H_t \hat{x}_{t|t-1})$ The term $(y_t - H_t \hat{x}_{t|t-1})$ is the innovation — the difference between what you observed and what you predicted. The Kalman gain scales this correction.

3. Update the covariance: $P_{t|t} = (I - K_t H_t) P_{t|t-1}$ Incorporating the observation always reduces uncertainty, so $P_{t|t}$ is smaller than $P_{t|t-1}$.

Forecasting and missing values:

• To forecast beyond the current time, recursively apply the prediction step without any update (since there are no future observations yet).
• To handle missing values, use the Kalman smoother, which runs a backward pass through the data after the forward filter. It uses future observations to refine past state estimates, effectively interpolating through gaps.

Performance evaluation of models

Accuracy metrics quantify how close your predictions are to actual values:

• MSE (Mean Squared Error) — average of squared prediction errors. Penalizes large errors heavily.
• RMSE (Root Mean Squared Error) — square root of MSE, so it's in the same units as your data. Easier to interpret than MSE.
• MAE (Mean Absolute Error) — average of absolute prediction errors. Less sensitive to outliers than MSE because it doesn't square the errors.
• MAPE (Mean Absolute Percentage Error) — expresses errors as percentages of actual values. Useful for comparing accuracy across different scales, but can be misleading when actual values are near zero.

Diagnostic tools check whether the model's assumptions hold:

• Residual analysis — the prediction errors (residuals) should be approximately normal, have constant variance (homoscedasticity), and show no systematic patterns over time.
• ACF and PACF of residuals — if the autocorrelation or partial autocorrelation functions show significant spikes, there's temporal structure the model hasn't captured.
• Q-Q plots and histograms — visual checks for whether residuals follow a normal distribution. Deviations in the tails of a Q-Q plot suggest the model underestimates extreme events.
• Ljung-Box test — a formal statistical test for autocorrelation in the residuals. A significant result (small p-value) means the residuals still contain predictable structure, indicating the model is inadequate.

State-space models vs other techniques

ARIMA works well for univariate series but requires stationarity (often achieved through differencing, which can lose information). It struggles with multiple related series and doesn't naturally accommodate missing data.

Exponential smoothing is simple to implement and effective for short-term forecasts where trend and seasonality are stable. It's limited when patterns are complex or time-varying.

Machine learning (neural networks, SVMs) can capture complex nonlinear relationships but requires large training datasets, is prone to overfitting without careful regularization, and produces models that are difficult to interpret in terms of underlying processes.

State-space models offer a middle ground: they're flexible enough to handle multivariate data, missing values, and time-varying dynamics while remaining interpretable. The main trade-off is that you need to specify the model structure (the $F_t$, $H_t$, $Q_t$, and $R_t$ matrices), which requires domain knowledge about the system you're modeling.