🤖 machine learning

Amortized Interventional Forecasting for Multivariate CIR Processes

This paper introduces CIR-ACTIVA, an amortized interventional forecasting framework for multivariate Cox--Ingersoll--Ross processes that enables causal "what-if" analysis of financial time series, such as CDS spreads, by distinguishing causal shocks from historical correlations and outperforming observational baselines in both causal selectivity and horizon-resolved calibration.

Original authors: Andreas Sauter, Sumit Sourabh, Drona Kandhai, Erman Acar

Published 2026-08-05

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Andreas Sauter, Sumit Sourabh, Drona Kandhai, Erman Acar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a flock of birds. Sometimes they all swoop left together, and sometimes they all dive right. If you just watch them, you might think they are all copying each other's moves. But what if one bird gets startled by a hawk and dives, causing the others to dive too? That's a cause-and-effect relationship. In the world of finance, money moves in similar flocks. Interest rates, credit scores, and the cost of borrowing money often rise and fall together. For a long time, computers used to predict these movements by just watching the patterns of the past, assuming that if two things moved together, they were linked. But this is like thinking the rooster causes the sun to rise just because the rooster crows every morning.

The specific tools scientists use to model these financial "flocks" are called CIR processes (named after three researchers: Cox, Ingersoll, and Ross). Think of these as a mathematical way to describe things that naturally want to return to a normal average, like a rubber band snapping back after being stretched. The big question this paper tackles is: What happens if we force one part of the system to change? If we artificially shock one company's credit score, how do the others react? Old models can't answer this because they only know how things have moved together, not how they would move if we pushed one of them. This paper introduces a new way to answer those "what-if" questions without having to rebuild the computer model every single time.

The Problem: The Rooster and the Sun

In the financial world, many things behave like a rubber band. If a credit spread (the cost of borrowing for a company) gets too high, it usually wants to snap back down to a normal level. If it gets too low, it wants to creep back up. This is called "mean-reverting."

For decades, scientists have used a standard model called the CIR process to predict these movements. It works great for guessing what might happen next based on history. But it has a blind spot. It sees that two things move together and assumes they are linked. It doesn't know why.

Imagine a stress test for a bank. A regulator might ask, "What happens if Company A's credit rating crashes?" A smart model should say, "Company B might get nervous and its rating might drop too, but Company C has nothing to do with Company A, so it should stay calm."

The old models, however, are like a confused observer who sees Company A and Company C moving together in the past. When Company A crashes, the old model panics and says, "Oh no, Company C must crash too!" It confuses correlation (things happening at the same time) with causation (one thing making the other happen). This leads to "phantom shocks," where the model predicts disasters for companies that are actually safe, causing banks to waste money on unnecessary insurance.

The Solution: A Time-Traveling Simulator

The authors of this paper, Andreas Sauter and his team, built a new tool called CIR-ACTIVA. They wanted to create a system that could answer "what-if" questions instantly, without needing to be retrained for every single new scenario.

To do this, they had to solve a tricky problem: How do you teach a computer about cause and effect if you can't actually go into the real world and crash a company's stock to see what happens? You can't run that experiment on real markets.

So, they built a virtual world. They created a "causal simulator" that generates fake financial data. In this fake world, they know exactly who causes whom. They can say, "Okay, in this simulation, Company A causes Company B to change, but Company C is totally independent." Then, they crash Company A in the simulation and watch what happens. This gives them the "ground truth"—the correct answer to the "what-if" question.

They trained their AI model, CIR-ACTIVA, on millions of these fake scenarios. The model learned to look at a history of movements and a specific "shock" (like "Force Company A to crash"), and then predict the future of the whole group.

The Secret Sauce: Time as a Character

Most AI models that predict the future treat time as a simple list of numbers. CIR-ACTIVA does something clever: it treats time itself as a character in the story.

Imagine you are telling a story about a ball rolling down a hill. If you just give someone a list of where the ball was, they might guess where it goes next. But if you tell them, "The ball was at the top at 1:00, halfway at 1:05, and at the bottom at 1:10," they understand the speed and the momentum.

CIR-ACTIVA looks at the data as a collection of snapshots, where every snapshot has a specific time stamp attached to it. It doesn't just learn "Company A goes down when Company B goes down." It learns "Company A goes down quickly right after the shock, but then slows down and settles over the next few days."

The model also uses a special trick called amortized inference. Usually, if you want to know what happens in a new scenario, you have to teach the computer from scratch. That takes a long time. CIR-ACTIVA is like a student who has studied so many different types of physics problems that when you give them a new problem, they can solve it instantly without opening a textbook. They have "amortized" (spread out) the learning across all the training scenarios.

The Results: Spotting the Real Victims

The team tested CIR-ACTIVA against other models, including some very smart ones that didn't understand time or causality. They used their virtual world to see who got the "what-if" answers right.

The results were clear. When they shocked one company in the simulation:

The Old Models (and the "GMR" baseline): They panicked. They predicted that everyone would crash, even the companies that had no connection to the victim. They couldn't tell the difference between a friend and a stranger.
CIR-ACTIVA: It stayed calm. It correctly predicted that the victim would crash, its direct friends would feel the pain, but the strangers would remain safe. It showed causal selectivity, meaning it only blamed the right people.

The model was also much better at predicting when things would happen. It knew that the shock would hit hard immediately but would fade away over time. The other models were often too slow or too fast in their predictions.

Why This Matters

This isn't just a math game. In the real world, if a bank thinks a safe company is in danger because of a "phantom shock," they might sell off their stock or buy expensive insurance they don't need. That costs real money.

The authors showed that their model works incredibly well in their simulated world. They even checked if their fake world looked like the real world, and while it wasn't perfect (real markets have more extreme jumps than their simulation), the model's ability to separate cause from effect held up.

The paper concludes that observational accuracy is not the same as causal correctness. Just because a model can guess the past well doesn't mean it can handle a shock. CIR-ACTIVA offers a way to run stress tests that actually tell you what would happen if you pushed a button, rather than just guessing based on what happened before. It's a step toward financial models that understand not just what is happening, but why.

Technical Summary: Amortized Interventional Forecasting for Multivariate CIR Processes

1. Problem Definition

Financial time series such as short rates, default intensities, and credit default swap (CDS) spreads are often modeled as mean-reverting processes, typically using the Cox–Ingersoll–Ross (CIR) framework. While standard multivariate CIR models successfully capture correlated co-movements, they lack a mechanism to represent causal influence between series. Consequently, they cannot answer interventional queries: predicting how a system responds when one series is externally shocked (e.g., a stress test forcing a specific spread to a stressed level).

Current forecasting models typically provide conditional forecasts (predicting future values given historical paths where a variable happened to take a certain level). This conflates correlation with causation. A high value in a correlated series may be a symptom of shared stress rather than a cause; thus, conditional models may incorrectly attribute shock effects to series that are causally shielded, leading to "phantom" risk assessments in stress testing.

The core challenge is interventional forecasting: estimating the post-intervention distribution $p(\mathbf{V} | \mathcal{D}_{M_o}, do(V_i))$ for multivariate CIR processes. This is complicated by observational equivalence, where multiple causal models generate identical observational data but imply different interventional outcomes. Furthermore, real-world financial data lacks ground-truth interventional samples, making evaluation of causal forecasting models impossible without synthetic benchmarks.

2. Methodology

The paper proposes CIR-ACTIVA, a framework combining a causal data generator and an amortized inference model.

2.1 Causal Multivariate CIR Simulator

To train and evaluate the model, the authors introduce a causal multivariate CIR data-generating process. This simulator generates paired observational and interventional samples, which are unavailable in real markets.

Structural Causal Model (SCM): The model defines $d$ variables where each evolves via a discretized CIR equation.
Causal Long-Run Mean: Unlike standard CIR models where the long-run mean is fixed, here the mean $b_{t+1}^i$ is a structural function of the contemporaneous values of causal parents ( $PaV_i$ ) and a time-modulation factor $c_i(t)$ .
Soft Interventions: The simulator applies soft interventions $do(V_i := (1+\delta)f_i)$ , rescaling the dynamics of a target variable while retaining its dependence on its parents.
Variants: The simulator includes a non-cyclical variant (stationary mean) and a cyclical variant where the long-run mean follows a sinusoidal modulation, testing the model's ability to track time-varying dynamics.
Calibration: Parameters are calibrated on real 5Y CDS data (2015–2021) to ensure marginal statistics (levels, autocorrelation, tail behavior) resemble real markets, though the simulator admits that real fat tails and shock persistence are harder to replicate with pure diffusion.

2.2 CIR-ACTIVA Model Architecture

The model adapts the ACTIVA framework (a $\beta$ -CVAE based amortized causal estimator) for time-series data.

Time-Series Framing: Instead of treating observations as an unordered set, the model promotes time to a variable. Each observation is a tuple $(\mathbf{y}, t)$ . This allows the model to learn temporal dependencies without imposing rigid architectural inductive biases (like RNNs), treating forecasting as a tabular prediction problem with time as a feature.
Encoder: Uses equivariant transformer encoders to process the observational context and intervention query. It employs a Conditional Continuous Normalizing Flow as the prior $p_\eta(\mathbf{z} | \mathcal{D}_{M_o}, \mathbf{i})$ to represent a multimodal latent distribution over causal models, preserving permutation equivariance.
Decoder: A transformer decoder maps the latent code $\mathbf{z}$ to a joint distribution over variables and time. Crucially, it uses a horizon-bucketed decoder. The forecast horizon is partitioned into $K$ time buckets, and the decoder emits a full-covariance Gaussian with a linear-in-time mean shift for each bucket. This allows the interventional law to vary with forecast distance.
Output Transform: To handle the non-negativity of CIR processes, the model predicts the square root of the spreads ( $\sqrt{x}$ ), stabilizing variance, and maps back to raw units via change of variables.
Training: The model is trained via a $\beta$ -CVAE objective with a large $\beta$ weight. This forces predictive structure into the conditioning path (observational data + intervention) rather than the latent code, ensuring the model can generalize to new tasks without retraining.

3. Key Contributions

CIR-ACTIVA Model: A novel amortized model for distributional causal effect estimation in multivariate CIR processes. It frames trajectories as time-stamped observations and utilizes a horizon-bucketed decoder to capture how shock responses evolve over different forecast horizons.
Causal CIR Simulator: A synthetic data-generating process that produces paired observational and interventional ground truth. This enables the training and rigorous evaluation of amortized causal inference models, a capability previously missing for time-series data.
Empirical Validation: The framework is instantiated on CDS spreads. The authors demonstrate that their approach outperforms baselines in both causal selectivity (shielding non-affected series from phantom shocks) and horizon-resolved calibration (accurately predicting the evolution of the shock response over time).

4. Experimental Results

The model was evaluated against four baselines: the original ACTIVA, MACE-TNP (a neural process estimator), a Gaussian Mixture Regression (GMR) baseline (non-causal), and an Oracle floor.

Causal Selectivity: CIR-ACTIVA significantly outperformed baselines in distinguishing between causal descendants and non-descendants.
- On non-descendants (causally shielded variables), the model's error (excess over Oracle) dropped by 51–56% relative to descendants.
- In contrast, baselines like GMR and ACTIVA failed to reduce error for non-descendants, often assigning "phantom" shock effects to them. MACE-TNP showed some selectivity but lost half its effectiveness when dynamics became time-varying (cyclical variant).
Horizon-Resolved Calibration: CIR-ACTIVA achieved the lowest error at short horizons, where the transition from observational to interventional dynamics is most critical.
- On cyclical data, the near-horizon Energy score was 0.46, significantly lower than MACE-TNP, ACTIVA, and GMR.
- The advantage persisted across all horizons on the non-cyclical dataset.
Ablation Study: Removing the expressive latent (flow prior) or the bucketed decoder degraded performance. The combination of both was essential for near-horizon accuracy, particularly in the cyclical setting.

5. Significance and Claims

The paper claims that observational forecasting accuracy is not a proxy for interventional correctness. In coupled financial systems, failing to distinguish correlation from causation leads to concrete errors: stress tests built on non-causal models book losses and size hedges against exposures that do not exist.

CIR-ACTIVA addresses this by:

Providing a method to answer "what-if" queries on coupled spread systems without retraining per scenario.
Establishing validity on synthetic ground truth that is independent of how well the simulator matches reality, while still demonstrating practical grounding through backtesting against real CDS data.
Offering a controlled benchmark (the causal CIR simulator) for future causal forecasters.

The authors are modest about the limitations: the synthetic ground truth understates shock persistence and the frequency of large moves (fat tails) found in real markets, meaning the reported scores likely represent an optimistic bound on real-world performance. Additionally, the current implementation is tested on systems of three names; scaling to portfolio-sized panels remains a separate challenge. However, the framework is agnostic to the specific asset class, applicable to any mean-reverting series modeled by CIR (e.g., short rates, stochastic volatility).