Time-Series Counterfactual Generation

The counterfactual must be minimally different from the original time series. This is typically enforced by an objective function that minimizes a distance metric, such as Dynamic Time Warping (DTW) or Euclidean distance, between the original instance x and the counterfactual x'. The goal is to identify the smallest possible intervention that flips the model's classification or forecast, ensuring the explanation is precise and not confounded by unnecessary alterations.

A valid counterfactual must remain within the distribution of plausible real-world data. Unconstrained optimization can produce adversarial examples that are physically impossible (e.g., a stock price jumping to a negative value). Techniques to enforce realism include:

• Autoencoder latent space projection: Decoding from a perturbed latent vector.
• Generative Adversarial Network (GAN) inpainting: Filling altered segments with realistic patterns.
• Shapelet constraints: Ensuring the counterfactual subsequence matches a known realistic motif.

To maximize interpretability, the counterfactual should alter the fewest possible time steps or features. A dense change across the entire sequence is difficult for a human operator to audit. Sparse counterfactuals are often achieved by adding an L1 regularization term to the loss function, which drives the difference vector (x' - x) to have many zero entries. This highlights a specific, localized temporal event as the root cause of the prediction.

The generated sequence must respect the autocorrelation structure and temporal dependencies of the original data. A counterfactual that introduces a sharp, instantaneous spike followed by an immediate return to baseline often violates the physics or inertia of the system (e.g., sensor telemetry, market volume). Constraints like maintaining the original power spectral density or using autoregressive models to generate the altered segment ensure the transition is smooth and physically coherent.

The altered features must correspond to variables a user can actually influence. In a financial context, a counterfactual suggesting 'change the market volatility index' is not actionable. A valid counterfactual for recourse translates model logic into a concrete intervention path, such as 'increase the liquidity reserve ratio by 2% at time step t-3.' This bridges the gap between statistical explanation and operational decision-making.

A purely associational counterfactual may suggest changing an effect to alter a cause, which is logically invalid. Advanced generation methods integrate a Structural Causal Model (SCM) of the time-series dynamics. The counterfactual is computed through a three-step process: abduction (infer the exogenous noise u), action (set a variable to a new value), and prediction (propagate the change forward through the causal graph). This guarantees the counterfactual trajectory respects the underlying causal mechanisms.