Counterfactual User Forecasting

• Counterfactual user behavior forecasting is the study of predicting how users' actions change under hypothetical interventions using structural causal models and deep learning.
• It employs methodologies like SCM, panel time series, and inverse propensity scoring to correct bias and simulate realistic action sequences.
• The approach delivers actionable insights for recommender systems and policy decisions by emphasizing interpretability, fairness, and robust causal evaluation.

Counterfactual User Behavior Forecasting is the discipline concerned with predicting how a user's future actions or behavioral trajectories would change under hypothetical interventions not actually observed in the historical data. This area integrates structural causal inference, time-series modeling, deep learning, and advanced evaluation criteria to provide interpretable and actionable "what-if" scenarios for decision support systems, recommender engines, and interactive online services. Research in this field addresses methodological challenges such as causal identifiability, bias correction under missing not-at-random exposure, modeling dependence on latent confounders, and generating realistic counterfactual sequences that satisfy business or process constraints.

1. Structural Formulation and Causal Graphs

Counterfactual user behavior forecasting requires a formal causal or counterfactual estimand that clarifies both the target intervention and the underlying system dependencies. Most state-of-the-art approaches use the potential outcomes framework (Rubin; Pearl), structural equation models (SEMs), or dynamic causal graphs.

Key paradigms include:

• Pearl-style SCMs: Nodes encode user states, platform features, exposures, intermediate adoption signals, and outcomes (e.g., "Counterfactual Forecasting of Human Behavior using Generative AI and Causal Graphs" (Uddandarao et al., 9 Nov 2025)).
• Panel and Time Series Models: Dynamic causal graphs or simultaneous graphical dynamic linear models (SGDLM) generalize dependency structure across multiple series, allowing explicit modeling of interventions on one or more behavioral streams ("Dynamic graphical models: Theory, structure and counterfactual forecasting" (West et al., 8 Oct 2024)).
• User and Item Decomposition: Disentanglement of user interest versus conformity, and item popularity versus intrinsic attributes (as in "Disentangled Counterfactual Reasoning" (Ren et al., 2023)), to enable precise targeting of the direct and indirect paths in user choice processes.

The counterfactual query is formalized as:

$P(Y_{t+1} \mid \text{do}(X_{t} = x'), C_{1:t})$

where $Y_{t+1}$ denotes the behavioral outcome, $X_{t}$ the manipulated exposure/intervention, and $C_{1:t}$ user covariates/history.

2. Methodologies for Counterfactual Forecasting

Several distinct methodologies have been proposed, each with specific modeling assumptions and application strengths:

Approach	Core Method	Typical Use Case
SCM + Transformer	Causal graph + generative	Scenario simulation for web/app/e-comm behavior
SGDLM (Bayesian)	Dynamic graphical models	Intervention effect in multivariate time series
Inverse Propensity	Likelihood reweighting	Bias correction, new-user event prediction
Doubly Robust	Cross-fitting with nuisance	Runtime confounding in personalized systems
Evolutionary Search	Sequence generation + Markov	Viable process analytics/trace counterfactuals
Simulation-based	SEM + RL-based intervention	Top-N ranking under hypothetical recommendations

Highlights:

• Gradient-based search: For time series forecasting, counterfactual histories are found via first-order optimization subject to forecast constraints (ForecastCF (Wang et al., 2023)).
• Multi-task learning: Simultaneous modeling of different aspects of sequential user interactions, e.g., click, conversion, and overall engagement (ESCIM (Ahn et al., 6 Oct 2025)).
• Contrastive/self-supervised: Exposure-aware contrastive sampling and InfoNCE losses enable deconfounding without explicit causal graphs ("Contrastive Counterfactual Learning" (Zhou et al., 2022)).
• Panel data factor models: Low-rank matrix completion, extended with factor dynamics, for "missing" counterfactual potential outcomes in longitudinal studies (FOCUS (Deb et al., 9 Nov 2025)).

3. Handling Bias, Confounding, and Exposure Mechanisms

Counterfactual user forecasting must rigorously adjust for selection, exposure, and confounding biases:

• IPS (Inverse Propensity Scoring): For unbalanced or MNAR exposure, instance weights invert the learned exposure probabilities, e.g.,

$$L_{IPS} = \sum_{(u,i):O_{u,i}=1} \frac{\delta(\hat{y}_{u,i},y_{u,i})}{P_{u,i}}$$

as in (Zhou et al., 2022) and sequential event forecasting for new users via IPW (Yuchi et al., 8 Jul 2024).

• Doubly-Robust Estimation: Combines propensity-score-corrected loss with outcome-model predictions to achieve consistency if either component is well specified ("Counterfactual Predictions under Runtime Confounding" (Coston et al., 2020); "Estimating and evaluating counterfactual prediction models" (Boyer et al., 2023)).
• Contrastive and Sampling Techniques: Random or propensity-guided counterfactual sampling expands the effective set of positive instances, simulating random exposure akin to RCTs (see Table 1 below).

Bias Correction	Evaluation/Effectiveness	Papers
IPS	Reduces MNAR bias, may increase variance	(Zhou et al., 2022, Yuchi et al., 8 Jul 2024)
Doubly Robust	Consistent under misspecification	(Coston et al., 2020, Boyer et al., 2023)
Contrastive Sampling	Enhances data efficiency, interpretable	(Zhou et al., 2022)

4. Generation and Evaluation of Counterfactual Sequences

A central challenge is generating not merely counterfactual scores, but entire sequences or trajectories that are both feasible and informative:

• ForecastCF (Wang et al., 2023) generates counterfactual time series histories $x_{cf}$ using gradient-based optimization of a constraint-masked loss, producing minimal, plausible perturbations that satisfy forecast bounds; validity and closeness are quantitatively evaluated.
• CREATED (Hundogan et al., 2023) employs evolutionary algorithms, with viability scored by (i) prediction delta, (ii) weighted edit similarity, (iii) sparsity, and (iv) process feasibility via a trained Markov model. This methodology maintains domain invariance and avoids infeasible counterfactuals.
• Panel Matrix Completion (FOCUS (Deb et al., 9 Nov 2025)) reconstructs missing potential outcomes for all units at all time points, then projects their future values via time-series dynamics on recovered latent factors.

Metrics combine validity (forecast falls within desired bounds), compactness (few changed points), and proximity (distance from factual sequence). Business constraints, such as seasonality or intervention feasibility, are incorporated as bound or edit constraints during counterfactual search.

5. Application Domains and Empirical Findings

Counterfactual user behavior forecasting frameworks are empirically validated across diverse domains:

• Conversion and recommendation: ESCIM (Ahn et al., 6 Oct 2025) improves both CVR and CTCVR AUCs by approx. +1% (offline), and yields +17.35% CVR gain (online) over strong baselines.
• Personalized recommendation: DCR (Ren et al., 2023) explicitly separates popularity and intrinsic user/item signals, removing bias from recommendation scores via direct-path interventions.
• Sequence modeling/LLMs: Counterfactual fine-tuning (CFT) (Zhang et al., 30 Oct 2024) augments transformer-based next-item forecasting, improving HR@K and NDCG@K by ∼9–10% across datasets.
• Panel/interventional studies: FOCUS (Deb et al., 9 Nov 2025) achieves up to 20–30% lower MSRPE than deterministic embeddings or "SyNBEATS" in real mHealth studies.
• Evaluations and model selection: DR estimators and counterfactual risk estimates allow honest tuning and ablation testing for real-world deployment environments (Boyer et al., 2023).

6. Challenges and Future Directions

Several open challenges and extensions are reported:

• Complex user confounding: Latent user types (e.g., category $C$ in (Yuchi et al., 8 Jul 2024)) or unmeasured time-varying confounders may remain unaddressed; ongoing work includes normalizing flows or variational methods for flexible propensity modeling.
• Scaling and computational efficiency: Edit-distance-based sequence comparison scales quadratically with sequence length; fast approximations or constraint learning are proposed (Hundogan et al., 2023).
• Causal graph learning and validation: Accurate causal structure determination with minimal domain knowledge remains difficult; hybrid data-driven and expert-in-the-loop methods are suggested (Uddandarao et al., 9 Nov 2025).
• Evaluation: There is no consensus on gold-standard metrics for counterfactual sequence validity and plausibility; most evaluations are empirical or rely on surrogate measures.
• Domain-specific adaptation: Extensions to include nonstationary dynamics, personalized covariate adjustment, and actionable intervention mapping are in active development.

7. Interpretability and Decision Support

Beyond pure forecasting accuracy, interpretability and actionable insights are primary motivations:

• Causal path visualization: Causal graphs learned by frameworks such as (Uddandarao et al., 9 Nov 2025) enable graphical trace of how interventions propagate through engagement and outcome layers; users can inspect quantitative impacts of direct and mediated paths.
• Actionable interventions: Outputs include mission-critical "what-if" analyses for A/B testing, simulated rollout, or algorithmic fairness audits—e.g., updating only feasible windows, enforcing monotonicity, or projecting predicted histories onto actionable policy sets (Wang et al., 2023, Deb et al., 9 Nov 2025).
• Debiasing and fairness: Removal of direct-path popularity/conformity effects (Ren et al., 2023) and use of DR estimators when deployment cannot match training conditions (Coston et al., 2020) allow for more equitable, robust user-outcome predictions.

In sum, counterfactual user behavior forecasting provides a principled, empirically validated, and increasingly versatile toolkit for simulating the effects of unobserved product, policy, or system interventions on user trajectories in complex, confounded, and dynamic digital environments.