💰 quantitative finance

PHINN: Persistent Homology Inspired Neural Network for Rare-Event Time Series Generation

PHINN is a flow-matching generative framework that leverages persistent homology, specifically dynamic Betti curves and persistence landscape losses, to effectively model rare events in time series by capturing stable topological fingerprints, thereby outperforming existing statistical and diffusion baselines in topological fidelity and tail coverage across diverse domains.

Original authors: Emre Yusuf, Ren Takahashi, Jayabrata Bhaduri

Published 2026-06-16✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Emre Yusuf, Ren Takahashi, Jayabrata Bhaduri

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster trying to predict a once-in-a-century hurricane. You have a lot of data about sunny days and light rain, but you've only seen one real hurricane in your entire career.

Most computer models try to learn from that single hurricane by looking at the numbers: "How fast was the wind? How low was the pressure?" But the paper argues that this is like trying to understand a hurricane just by looking at a spreadsheet of wind speeds. You miss the shape of the storm—how the eye forms, how the clouds swirl in a specific loop, and how the system connects together.

This paper introduces PHINN (Persistent Homology Inspired Neural Network), a new kind of AI designed to invent realistic "what-if" scenarios for rare, disastrous events (like financial crashes, supply chain failures, or cyber-attacks) by focusing on their geometric shape rather than just their statistics.

Here is how it works, broken down into simple concepts:

1. The "Shape" of a Crisis (Topology)

The authors realized that rare events leave a unique "fingerprint" in the data, not just in the numbers, but in the structure.

The Analogy: Imagine looking at a crowd of people.
- Statistics tell you the average height and weight.
- Topology (the math used here) tells you the shape of the crowd. Are they all standing in one big circle? Are there three separate groups chatting? Is there a hole in the middle where no one is standing?
The Insight: A stock market crash looks "topologically" different from a supply chain blockade, even if the numbers (price drops) look similar. PHINN learns to recognize these shapes. It counts "loops" (like a ring of connected events) and "voids" (empty spaces in the data) to understand the crisis.

2. The "Dynamic Camera" (Sliding Windows)

Time series data is a movie, not a photo. PHINN doesn't look at the whole movie at once.

The Analogy: Imagine a security camera that zooms in and out automatically. When things are calm, the camera zooms out to see the big picture. When chaos starts (a "crisis"), the camera zooms in tight to catch the fast-moving details.
How PHINN does it: It uses a "sliding window" that changes size based on how volatile the data is. It turns the time series into a cloud of points and constantly checks: "How many separate groups are there? How many loops are forming?" These changing counts are called Betti curves.

3. The "Architect" (Flow Matching)

Once PHINN understands the "shape" of a crisis, it needs to generate new, fake-but-realistic scenarios.

The Analogy: Think of a sculptor who has a block of clay (random noise). A normal AI might just squish the clay randomly. PHINN is a sculptor who has a blueprint (the Betti curve). It forces the clay to take on the specific shape of a "market crash" or a "cyber-attack" as it molds it.
The Result: It can generate 10,000 different versions of a disaster scenario that all look structurally correct, helping planners prepare for things they haven't seen before.

4. The "Translator" (LLM Interface)

You don't need to be a mathematician to use this.

The Analogy: You can talk to PHINN in plain English. You can say, "Show me a supply chain failure where a key factory shuts down for two weeks."
How it works: A language model (like a smart chatbot) translates your sentence into the specific "shape blueprint" (Betti curve) that the AI needs to build the scenario.

5. Why This Matters (The "What If" Machine)

The paper claims that current methods fail to capture the structure of rare events. They might get the numbers right but the story wrong.

The Claim: PHINN is better at creating scenarios that look and feel like real disasters. It is more accurate at predicting the "shape" of the crisis (up to 63% more accurate in some tests) and matches the performance of expert human planners in creating realistic "tail" events (the worst-case scenarios).
The Safety Check: The authors also built a "lie detector" into the system. It can spot if someone tries to trick the AI with fake data that looks statistically okay but has the wrong "shape."

Summary

PHINN is a tool that helps us imagine the unimaginable. Instead of just crunching numbers to guess what a rare disaster might look like, it learns the geometric DNA of past disasters. It then uses that knowledge to build thousands of new, realistic "what-if" stories, helping companies and governments prepare for the next big shock before it happens.

Key Takeaway: It's not just about how bad the numbers get; it's about how the chaos is shaped. PHINN learns the shape of the storm so it can predict the next one.

Technical Summary: PHINN – Persistent Homology Inspired Neural Network for Outlier Synthesis

1. Problem Statement

The paper addresses the critical challenge of modeling rare, high-impact events in time series data, such as financial crises, infrastructure failures, and cyber-physical anomalies. Existing approaches face two primary limitations:

Data Scarcity: Rare events are too infrequent to learn from effectively using standard supervised or generative methods.
Structural Blindness: Current generative models (including diffusion models) struggle to capture the structural dynamics of extreme values. While statistical methods (e.g., Extreme Value Theory, heavier tails) address marginal distributions, they fail to model the geometric shape of a crisis. Recent surveys indicate diffusion models "universally struggle with extreme values," and practitioners lack tools to generate structurally faithful scenarios from a single observed example.

The core operational gap identified is the inability to answer: "What does a disruption of this structural type look like dynamically, and what are plausible ways it unfolds?"

2. Methodology: PHINN

The authors propose PHINN, a conditional flow-matching framework designed to synthesize rare-event time series by leveraging topological data analysis (TDA).

Core Insight

The authors observe that rare events leave distinct topological fingerprints. When a time series is embedded into a sliding-window point cloud, the evolution of its Betti numbers ( $\beta_0, \beta_1, \beta_2$ ) reveals characteristic transitions (e.g., changes in connected components, loops, or voids) that are more stable and discriminative than statistical moments. These transitions often precede or coincide with statistical shocks.

Architecture and Components

PHINN integrates four key technical modules:

Topological Fingerprinting Engine:
- Converts time series into sliding-window point clouds using Takens' embedding.
- Computes persistent homology (Vietoris–Rips filtration) to extract Betti curves $\beta(t) = (\beta_0(t), \beta_1(t), \beta_2(t), \chi(t))$ .
- Implements dynamic window sizing, where the window width contracts during high-volatility periods and expands during calm, adapting to local total variation.
- Updates are handled efficiently via union-find structures, achieving low latency ( $\le 98$ ms) on streaming data.
Differentiable Topological Loss:
- To enforce topological consistency during training, the model uses a Persistence Landscape Loss ( $L_{topo}$ ).
- This loss operates on $H_1$ and $H_2$ homology groups (extending prior $H_1$ -only work) using smoothed Gaussian-kernel approximations to ensure differentiability.
- The total objective combines Rectified Flow Matching loss ( $L_{RM}$ ), topological loss, and statistical moment matching ( $L_{stat}$ ).
Topology-Conditioned Flow Matching:
- The model employs a rectified flow ODE where the velocity field $v_\theta$ is conditioned on the target Betti curve.
- A Transformer encoder processes the Betti curve (with positional encoding) to generate a conditioning vector $c$ , which is integrated via cross-attention.
- Inference is deterministic and fast ( $<500$ ms).
Multivariate and Cross-Domain Extensions:
- Joint Topology: For multi-modal data, PHINN constructs a joint point cloud from heterogeneous sensors and applies joint Vietoris–Rips filtrations to detect topological entanglement between modalities.
- LLM Interface: A fine-tuned LLM (Mistral-7B) translates natural language scenario descriptions into specific Betti-curve conditioning targets.
- Meta-Learning: Uses MAML-style training to enable 1-shot fine-tuning across heterogeneous rare-event corpora.
- Retrieval-Augmented Generation: A topological memory bank allows for few-shot generation by retrieving similar Betti curves and blending them with the query.

Robustness

The framework includes certified adversarial robustness:

Type-I: Formal $\ell_\infty$ certificates based on the stability of persistence diagrams under bounded noise.
Type-II: An empirical mechanism to detect structurally coherent adversarial injections by comparing raw statistics against expected topological consistency.

3. Key Contributions

The paper claims the following specific contributions:

Dynamic Betti-Curve Conditioning: The first use of time-varying Betti curves as continuous conditioning signals for time-series generation, featuring volatility-adaptive window sizing.
Differentiable Higher-Order Homology Loss: A persistence landscape formulation that extends topological losses to $H_1$ and $H_2$ , enabling end-to-end training.
Topology-Conditioned Flow Matching: A rectified flow ODE architecture with Betti cross-attention, achieving sub-second inference.
Multivariate Joint Topology: A method for joint Vietoris–Rips filtrations on product point clouds to capture cross-modal topological coupling.
Natural Language Interface: A learned translation layer mapping practitioner intent (natural language) to Betti-curve targets.
Cross-Domain Meta-Learning: Demonstrated 1-shot fine-tuning capabilities with 5 $\times$ data efficiency across diverse domains.
Certified Robustness: Formal guarantees for persistence diagram stability and empirical detection of structural adversaries.
Topological Scenario Atlas: A validated dictionary mapping specific Betti transitions (e.g., $\beta_0: 1 \to 3$ ) to operational decision triggers.

4. Experimental Results

The authors evaluated PHINN on financial crises, epidemiological data, and multi-modal time series (157 labeled rare events).

Topological Fidelity: PHINN significantly outperformed statistical and diffusion-based baselines (including FIDE, EFDiff, and FM-TS).
- $\beta$ -RMSE: Reduced by 41–63% compared to baselines.
- Transition Accuracy: Increased by 84%.
Statistical Fidelity: PHINN matched practitioner-grade jump-diffusion models in tail coverage (ECov95) while substantially exceeding them in structural shape fidelity.
Ablation: The unconditional version of PHINN (PHINN-Uncond) performed statistically indistinguishably from standard flow matching (FM-TS), confirming that the performance gains are driven by topological conditioning rather than architectural capacity.
Operational Value: In a decision study with domain experts, PHINN-generated scenarios led to higher Decision Change Rates (DCR) compared to human-generated or FIDE scenarios, indicating better alignment with contingency planning needs.
Robustness: The system achieved 84% recall in detecting structural adversarial attacks at a 1% false alarm rate.

5. Significance and Claims

The paper positions PHINN as a novel intersection of topology-conditioned generation, flow matching, and rare-event time-series synthesis.

Novelty: The authors claim no prior work has combined dynamic Betti-curve conditioning with flow matching for outlier synthesis. While topology-conditioned models exist for static data (3D shapes, 2D masks), PHINN is the first to apply this to dynamic time series.
Interpretability: By using Betti numbers, the model provides an interpretable geometric structure for rare events, bridging the gap between abstract kernel embeddings and actionable scenario generation.
Practical Utility: The "Topological Scenario Atlas" provides a validated mapping between topological transitions and operational decisions (e.g., "Activate dual-source protocol" upon $\beta_0: 1 \to 3$ ), moving beyond mere statistical simulation to structurally faithful scenario planning.

The authors conclude that PHINN offers a robust, interpretable, and efficient framework for generating plausible, structurally accurate rare-event scenarios, addressing a critical gap in risk management and contingency planning.