💻 computer science

Haar-CFR: Solving Imperfect-Information Games with Applications in Data Analytics and Finance

Haar-CFR is a novel hybrid framework that integrates discrete Haar wavelet regularization with the CFR+ algorithm to achieve massive memory compression and faster convergence than existing methods, offering a scalable and interpretable solution for solving large-scale imperfect-information games in domains like finance and poker.

Original authors: Florian Fey, Sergei Gorlatch

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Florian Fey, Sergei Gorlatch

Original paper licensed under CC BY 4.0 (https://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

In the complex world of high-stakes decision-making, from the card tables of professional poker to the lightning-fast exchanges of global finance, there is a persistent struggle between two opposing forces: the need for perfect memory and the need for instant speed. Imagine a market maker, a digital agent whose job is to buy and sell assets to keep the market flowing. This agent must constantly balance two conflicting dangers. First, there is the risk of holding too much of an asset, which forces them to lower prices to get rid of it. Second, there is the risk of trading with someone who knows more than they do, which forces them to widen the gap between buying and selling prices to protect themselves. To find the perfect strategy, computers traditionally try to calculate every possible future scenario. However, as the number of possibilities grows, the computer's memory fills up and crashes, a problem known as the "memory wall." Alternatively, researchers have tried using artificial intelligence to guess the answers, but these systems are often slow to think and difficult to understand, creating a "time wall" that prevents them from acting fast enough in real-time.

A team of researchers at the University of Münster has developed a new method called Haar-CFR that bridges this gap, allowing computers to solve these massive decision problems without running out of memory or getting stuck in slow calculations. Instead of trying to store every single detail of every possible situation in a giant, rigid list, the researchers realized that the patterns of regret—the feeling of having made a wrong choice—follow a structured shape. They treated these patterns like a signal, similar to how a sound wave can be broken down into its essential frequencies. By using a mathematical tool called a wavelet transform, they could compress the massive amount of data into a tiny, efficient summary that kept all the important strategic information while discarding the noise. This approach allowed their system to handle games with millions of possible states using a fraction of the memory required by traditional methods, while still running fast enough to make decisions in real-time.

The researchers tested this new system on several challenging benchmarks, including complex variations of poker and a sophisticated simulation of financial market making. In the poker tests, they pushed the system to handle a game state space that was ten thousand times larger than what standard computer memory could usually manage. While traditional methods ran out of memory and stopped working, the new system continued to operate smoothly, using less than two megabytes of storage even when the game size grew to ten million possibilities. This represents a reduction in storage needs of more than 50,000 times compared to the old way of doing things. Furthermore, when compared to modern artificial intelligence systems that rely on deep neural networks, the new method was six and a half times faster in actual clock time. This speed advantage came because the system avoided the heavy computational cost of running complex neural network calculations, relying instead on fast, direct mathematical operations that a standard computer processor could handle instantly.

In the financial simulation, the system was tasked with solving the "impossible triangle" of market making, where an agent must simultaneously manage inventory risks and protect against informed traders. Previous theories had treated these two risks as separate problems, with one set of rules for managing stock levels and another for handling information leaks. The new system, however, learned to unify these conflicting pressures into a single, coherent strategy. It discovered a non-linear pattern where the agent would act normally when the market was calm but would suddenly switch to a defensive mode when the risk of trading with a better-informed opponent became too high. In these high-risk moments, the agent effectively stopped trading to avoid losses, a behavior that emerged naturally from the calculation rather than being programmed in advance. This result showed that the system could recover the logic of classical financial theories while also adapting to complex, unpredictable market conditions.

One of the most significant findings of this work is that the system is not a "black box" where decisions happen in the dark. Because the method uses a specific mathematical structure to compress the data, researchers can look inside the system and see exactly which factors are driving the decisions. By analyzing the compressed data, they could visually separate the agent's reaction to inventory levels from its reaction to market toxicity. This transparency means that regulators or human supervisors could audit the system to ensure it is making safe and logical choices, a feature that is often missing in modern artificial intelligence. The researchers found that the system's ability to learn depended on keeping a specific number of these compressed data points; if they kept too few, the system failed to learn the sharp boundaries between different strategies, but if they kept enough, it achieved results that were statistically indistinguishable from the perfect, memory-heavy solutions.

The study suggests that treating complex decision problems as signals that can be compressed is a powerful alternative to both brute-force memory storage and slow neural network guessing. The researchers demonstrated that for many real-world problems, the underlying strategies are not random chaos but structured patterns that can be represented very efficiently. While the method works best when the strategies have these clear, structured patterns, it offers a robust and scalable path forward for solving large-scale games and financial models. By combining the rigorous guarantees of game theory with the efficiency of signal processing, the team has provided a new tool that is both fast and understandable, potentially changing how computers approach complex strategic challenges in economics and data science.

Technical Summary: Haar-CFR

1. Problem Statement

The paper addresses the fundamental challenge of solving large-scale imperfect-information games, a core task in data analytics, machine learning, and decision-making. Current state-of-the-art approaches face a "double bottleneck":

The Memory Wall (Tabular Methods): Classical Counterfactual Regret Minimization (CFR) and its accelerated variant, CFR+, guarantee convergence to an $\epsilon$ -Nash equilibrium. However, they rely on dense tabular storage for every information set. In realistic environments with fine-grained discretizations (e.g., continuous price grids or large hand spaces), this leads to combinatorial state space explosion, rendering the methods intractable on standard hardware.
The Time Wall (Neural Methods): Deep Reinforcement Learning (DRL) and Deep CFR attempt to mitigate memory issues by approximating regret surfaces with neural networks. While flexible, these methods introduce prohibitive latency due to neural inference overhead, preventing deep lookahead in real-time applications. Furthermore, they suffer from training instability, simulator overfitting, and a lack of interpretability.

The authors identify a specific application stress test for these limitations: Automated Market Making. This domain requires balancing conflicting objectives—Inventory Risk (needing price adjustments to liquidate positions) and Adverse Selection Risk (needing spread widening to defend against informed traders). Existing theoretical models (e.g., Avellaneda–Stoikov for inventory, Glosten–Milgrom for information) are often disjoint and analytically intractable when unified, while computational approaches struggle with the required scalability and interpretability.

2. Methodology: Haar-CFR

The authors propose Haar-CFR, a hybrid soft computing framework that integrates Discrete Haar Wavelet transforms with CFR+ to resolve the storage bottleneck without incurring the inference latency of neural networks.

Core Hypothesis: The Sparse-Haar Hypothesis

The method relies on the observation that regret landscapes over ordered domains (such as price grids or hand strengths) are not random noise but structured signals. Specifically, equilibrium strategies in these domains often exhibit regime-switching behavior, remaining constant over intervals and changing only at specific thresholds.

Hypothesis: The equilibrium regret function is piecewise-constant with a limited number of jump discontinuities ( $S$ ). Consequently, its Discrete Haar Wavelet expansion contains at most $O(S \log N)$ non-zero coefficients, where $N$ is the grid size. This implies the regret signal is highly compressible.

Algorithmic Workflow

Haar-CFR replaces the dense tabular storage of standard CFR with a Sparse Spectral Registry:

Reconstruction (Transient): During node traversal, the algorithm retrieves sparse wavelet coefficients from persistent storage and reconstructs the dense regret vector using the Inverse Discrete Haar Transform (IDHT).
Strategy Computation: The dense vector is used to compute the local policy via Regret Matching+ (as in CFR+).
Accumulation: Instantaneous counterfactual regrets are accumulated to update the dense vector.
Spectral Regularization (Compression): The updated dense vector is transformed back into the spectral domain using the Forward Discrete Haar Transform (DHT). An adaptive Top-K thresholding step retains only the largest $K$ coefficients, discarding the rest.
Storage: Only the sparse coefficients are committed to persistent memory.

This process ensures that memory usage scales near-logarithmically with the state space size ( $O(\log N)$ ) rather than linearly ( $O(N)$ ), while the dense vector exists only transiently in RAM.

Theoretical Guarantees

The paper provides a theoretical analysis showing that if the truncation error (due to compression) grows sublinearly over iterations, the "no-regret" property of CFR is preserved. Under the Sparse-Haar hypothesis, the approximation error decays sufficiently fast to ensure convergence to an $\epsilon$ -Nash equilibrium, provided the number of retained coefficients $K$ satisfies the information-theoretic lower bound ( $K \geq S \log |C|$ ). If $K$ falls below this limit, the basis functions become insufficient to represent strategic discontinuities, leading to convergence collapse.

3. Key Contributions

Algorithmic Innovation: Introduction of Haar-CFR, a spectral regularization algorithm that achieves compression ratios exceeding 99.9% (reducing storage by $>50,000\times$ ) when the compression schedule maintains the required coefficient count ( $K \geq S \log |C|$ ), preserving equilibrium convergence.
Efficiency: Demonstration that Haar-CFR overcomes the "Time Wall" of neural methods. By avoiding neural inference and utilizing fast CPU-based Haar transforms, it achieves a 6.5× wall-clock speedup over tuned Deep CFR baselines on high-dimensional domains.
Unified Financial Modeling: Successful application to a high-dimensional market making environment where the agent autonomously unifies distinct theoretical paradigms (Inventory Risk and Adverse Selection) into a single equilibrium model.
Explainability: Provision of intrinsic explainability through spectral decomposition. Unlike black-box neural networks, Haar-CFR allows for the visualization of specific strategic frequencies (e.g., separating inventory management signals from adverse selection signals).

4. Experimental Results

The framework was evaluated on three standard game benchmarks (Leduc Hold'em, Deep-Stack Kuhn Poker, Flop Hold'em) and a custom financial market making stress test.

Scalability (Memory): In the Deep-Stack Kuhn Poker benchmark (scaling from $10^3$ to $10^7$ states), standard Tabular CFR+ exhausted memory (OOM) at $N=10^5$ . Haar-CFR maintained a memory footprint of only 2.1 MB at $N=10^7$ , demonstrating near-logarithmic scaling.
Speed (Time): On the Flop Hold'em benchmark, Haar-CFR reached a target exploitability of 100 mbb/h in approximately 4 hours, compared to 26 hours for Deep CFR, representing a 6.5× speedup.
Robustness: Ablation studies confirmed that retaining the top 5% of coefficients ( $\kappa=0.05$ ) yields near-lossless abstraction. However, aggressive compression ( $\kappa=0.001$ ) led to convergence failure, validating the theoretical lower bound on the number of required coefficients.
Financial Application: In the market making domain, the learned strategy recovered the linear inventory skew of Avellaneda–Stoikov at low toxicity and the spread widening of Glosten–Milgrom at high toxicity. Crucially, it discovered a Strategic Hybrid Sigmoidal Policy, a non-linear surface that dynamically switches regimes, effectively creating a "Defensive Wall" against toxic flow.
Statistical Superiority: Haar-CFR significantly outperformed Deep CFR in the financial domain ( $p < 0.001$ ), attributed to the "Spectral Bias" of neural networks which struggle to approximate high-frequency discontinuities (sharp decision boundaries) without excessive training.

5. Significance and Claims

The paper claims that Haar-CFR establishes wavelet-regularized regret minimization as a novel, broadly applicable, and interpretable alternative to both tabular and neural equilibrium solvers.

Bridging the Gap: It offers a "third path" that avoids the memory explosion of exact tabular methods and the latency/interpretability issues of neural approximations.
Signal Processing in Game Theory: By treating the regret landscape as a signal rather than a table, the method leverages the sparsity of strategic thresholds to achieve high-fidelity modeling in structured, imperfect-information games.
Explainable AI (XAI): The framework provides a rigorous mechanism for auditing algorithmic decision-making in safety-critical domains (like finance) by decomposing policies into orthogonal spectral components, revealing exactly which risk factors drive specific behaviors.

The authors conclude that while the method is bounded by the sparsity of the equilibrium strategy (failing if the strategy is highly oscillatory or fractal, or if the coefficient count falls below the information-theoretic limit), it successfully solves previously intractable high-dimensional problems in economics and data science.