DoFlow: Flow-based Generative Models for Interventional and Counterfactual Forecasting on Time Series

DoFlow is a flow-based generative model built on causal Directed Acyclic Graphs that unifies observational forecasting, interventional predictions, and counterfactual reasoning for multivariate time series while enabling principled anomaly detection through explicit likelihood estimation.

The Gist

Imagine you are a weather forecaster. Usually, you look at the past few days of rain and wind to predict if it will rain tomorrow. That's observational forecasting: "Based on what happened, what will likely happen next?"

But what if you want to answer deeper questions?

• Interventional: "What if I build a giant dam upstream? How will the river flow change downstream?"
• Counterfactual: "It rained heavily yesterday and the crops died. But what if I had built that dam yesterday? Would the crops have survived?"

Most modern AI models are great at the first question (predicting the future based on the past) but terrible at the second and third. They are like a car that can drive forward but has no reverse gear and no steering wheel to change direction.

This paper introduces DoFlow, a new AI model designed to be the ultimate "Time Traveler's Dashboard." It doesn't just predict the future; it lets you simulate "What If" scenarios and even rewind time to see how things could have been different.

The Core Idea: The Causal Map

To understand DoFlow, imagine a complex machine, like a hydropower plant or a human body.

• In a normal AI, all the parts are just a big soup of numbers. If the turbine spins fast, the generator hums. The AI learns this pattern.
• In DoFlow, the AI is given a Causal Map (a Directed Acyclic Graph, or DAG). Think of this as a plumbing diagram. It knows that Water $\to$ Turbine $\to$ Generator $\to$ Power Grid. It understands that the Turbine causes the Generator to spin, not the other way around.

How It Works: The "Flow" Machine

The magic behind DoFlow is something called a Continuous Normalizing Flow (CNF). Let's use a metaphor to explain this:

Imagine you have a lump of clay (the data).

1. The Encoder (The Sculptor): DoFlow looks at the current state of the system and "sculpts" it down into a simple, smooth ball of clay (a standard mathematical shape called a Gaussian distribution). This is like compressing a complex movie into a single, perfect file.
2. The Flow (The Conveyor Belt): Because this process is a "flow," it is perfectly reversible. We can run the conveyor belt backward.
3. The Decoder (The Builder): To predict the future, DoFlow takes a fresh, random ball of clay and runs it through the conveyor belt in reverse, turning it back into a complex shape (a prediction of the future).

Why is this special?
Because the process is reversible and mathematically precise, DoFlow can do two things normal models can't:

• Intervention: If you want to see what happens if you force the Turbine to stop, you just stop the clay at that step in the process. The rest of the machine automatically recalculates how the Generator and Grid will react.
• Counterfactual: If you want to know "What if the Turbine had stopped yesterday?", DoFlow can take the actual clay ball from yesterday, rewind it to the "Turbine Stopped" state, and then run it forward again to see the new timeline.

Real-World Superpowers

The paper tests DoFlow on two very different worlds:

1. The Hydropower Plant (The "What If" Safety Check)
Imagine a dam. Sensors measure water flow, vibration, and electricity.

• The Problem: Sometimes a turbine breaks. By the time the power grid sees the failure, it's too late.
• DoFlow's Job: It watches the sensors. If the vibration patterns start to look "weird" (even before a total breakdown), DoFlow calculates the probability of the future. If the probability of a "normal" future drops too low, it sounds an alarm 20 minutes before the actual failure. It's like a doctor who can tell you are getting sick before you even feel a fever.

2. Cancer Treatment (The "What If" Medical Trial)
Imagine a patient undergoing chemotherapy.

• The Problem: Doctors need to know: "If we gave this patient a higher dose of radiation, would they have done better?" We can't go back in time to test this on the real patient.
• DoFlow's Job: It takes the patient's actual history, "rewinds" the timeline to the point where the dose was different, and simulates the new outcome. It acts as a virtual time machine for doctors to test treatments without risking the patient's life.

Why This Matters

Currently, if you want to know the effect of a new policy, a new drug, or a new machine setting, you often have to wait for it to happen in the real world, or run expensive, risky experiments.

DoFlow allows us to:

• Predict the future (Observational).
• Simulate changes (Interventional).
• Rewrite history to learn from it (Counterfactual).

It unifies the ability to "guess what happens next" with the ability to "ask what if," all while understanding the cause-and-effect rules of the universe. It's a step toward AI that doesn't just see patterns, but truly understands how the world works.

Technical Summary

1. Problem Statement

Time-series forecasting has traditionally focused on observational prediction: estimating future values based on historical correlations. However, many real-world applications (e.g., healthcare, energy management) require answering causal queries:

• Interventional Queries: "What will happen if we change specific control variables (e.g., turbine settings, drug dosages)?"
• Counterfactual Queries: "What would have happened to this specific patient's trajectory if they had received a different treatment, given their observed history?"

Limitations of Existing Methods:

• Standard deep learning models (Transformers, RNNs) are purely observational and cannot simulate the effects of interventions or generate counterfactuals.
• Existing causal methods often focus on static data or discrete, fixed-time actions, lacking a general framework for continuous, time-indexed causal forecasting on Directed Acyclic Graphs (DAGs).
• There is a lack of models that can simultaneously provide accurate observational forecasts, coherent interventional predictions, and counterfactual trajectories with explicit likelihoods.

2. Methodology: DoFlow

The authors propose DoFlow, a generative framework that unifies causal reasoning with deep generative modeling using Continuous Normalizing Flows (CNFs) defined over a causal DAG.

Core Architecture

• Causal Structure: The system is modeled as a multivariate time series evolving over a known causal DAG. The value of node $i$ at time $t$ ($X_{i,t}$) depends on its own history, the history of its parents in the DAG, and exogenous noise.
• Time-Conditioned CNF:
  • Encoding (Abduction): For each node, an RNN (LSTM/GRU) summarizes the past trajectories of the node and its parents into a hidden state $H_{i,t-1}$. A CNF encoder maps the observed factual value $x_{i,t}$ to a latent representation $z_{i,t}$ conditioned on $H_{i,t-1}$.
  • Velocity Field: The transformation from the base distribution (Gaussian noise) to the data distribution is governed by a Neural Ordinary Differential Equation (Neural ODE). The velocity field $v_\theta$ is trained using Conditional Flow Matching (CFM), which minimizes the $L_2$ distance between the learned velocity and an analytically defined reference velocity (linear interpolation between data and noise).
  • Decoding (Prediction):
    • Observational/Interventional: The model samples from the base distribution $z \sim \mathcal{N}(0, I)$ and integrates the Neural ODE backward (reverse process) to generate future trajectories. Interventions are handled by fixing specific nodes to intervention values ($\gamma_{i,t}$) during the topological generation process.
    • Counterfactual: The model follows the Abduction-Action-Prediction paradigm:
      1. Abduction: Encode the observed factual trajectory into latent variables $z^F$ using the factual hidden states.
      2. Action: Apply the hypothetical intervention schedule.
      3. Prediction: Decode the same latent variables $z^F$ using the counterfactual hidden states (updated with the intervention) to generate the counterfactual trajectory.

Theoretical Guarantees

The paper provides a Counterfactual Recovery Theorem (Corollary 4.5). Under assumptions of:

1. Independence of exogenous noise across nodes/time.
2. Strict monotonicity of the structural causal mechanism with respect to noise.
3. The encoded latent variable being statistically independent of the conditioning history (a property of exact CNF training).

The authors prove that DoFlow can almost surely recover the true counterfactual trajectory. The latent variable $Z_t$ acts as a function of the exogenous noise $U_t$ that is invariant to the specific history, allowing the model to "replay" the same noise under a different causal regime.

3. Key Contributions

1. Unified Framework: DoFlow is the first general framework capable of performing observational, interventional, and counterfactual forecasting on time-series data structured by a causal DAG.
2. Counterfactual Generation: It enables the generation of coherent counterfactual trajectories by leveraging the invertibility of CNFs to preserve the underlying exogenous noise (abduction) while changing the causal mechanism (intervention).
3. Likelihood-Based Anomaly Detection: Unlike standard point-forecasters, DoFlow provides explicit log-likelihoods for future trajectories. This allows for principled anomaly detection by identifying trajectories with low probability density under the learned dynamics.
4. Theoretical Recovery: The paper establishes theoretical conditions under which the model perfectly recovers counterfactuals, bridging the gap between generative modeling and causal inference theory.

4. Experimental Results

The authors evaluated DoFlow on synthetic datasets with various DAG structures (Tree, Diamond, FC-Layer, Chain) and real-world applications.

• Synthetic Data:

  • Metrics: Root Mean Squared Error (RMSE), Maximum Mean Discrepancy (MMD), and Continuous Ranked Probability Score (CRPS).
  • Performance: DoFlow consistently outperformed strong baselines (GRU, TFT, TiDE, DeepVAR, MQF2) in both observational and interventional forecasting.
  • Counterfactuals: DoFlow was the only method capable of generating counterfactual trajectories. Baselines failed to produce meaningful counterfactuals (marked as NA in tables).
  • Robustness: The model performed well on both additive noise models and complex non-linear, non-additive (NLNA) structural causal models.

• Real-World Applications:

  • Hydropower System: Used data from Argonne National Laboratory. DoFlow successfully predicted system-wide trajectories following turbine failures (interventions) and detected power outages 10–20 minutes in advance using log-density scoring, outperforming all baselines in anomaly detection.
  • Cancer Treatment: Evaluated on the Bica et al. (2020) dataset for estimating treatment effects. DoFlow achieved significantly lower Normalized RMSE compared to CRN, RMSN, and MSM for predicting tumor volumes under different chemotherapy/radiotherapy dosing schedules.

5. Significance and Impact

• Decision Support: DoFlow moves beyond correlation-based prediction to enable "what-if" analysis, which is critical for high-stakes domains like healthcare (treatment planning) and engineering (system control).
• Trustworthy AI: By providing explicit likelihoods and theoretically grounded counterfactuals, the model offers a more interpretable and reliable tool for decision-making in complex dynamical systems.
• Future Directions: The work lays the foundation for integrating causal discovery with generative modeling, potentially allowing the system to learn the DAG structure directly from data in future iterations.

In summary, DoFlow represents a significant step forward in unifying causal inference and deep generative modeling, offering a robust solution for forecasting under interventions and generating counterfactual scenarios in multivariate time-series systems.