torchfbm 0.1.2

High-performance Fractional Brownian Motion toolkit for PyTorch with generators, processes, neural layers, and RL support

torch-fbm

Differentiable Fractional Brownian Motion & Rough Volatility for PyTorch

torch-fbm is a high-performance, GPU-accelerated library for generating and analyzing Fractional Brownian Motion (fBm) and Fractional Gaussian Noise (fGn).

Designed for Quantitative Finance (Rough Volatility, Real-Time Streaming), Deep Reinforcement Learning (Regime-Aware Exploration), and Generative Modeling (Rough Diffusion), it provides differentiable generators and layers that seamlessly integrate into the PyTorch ecosystem.

---

Features

Core Generators

• Fast Generation: Davies–Harte algorithm (FFT-based) for $O(N \log N)$ complexity.
• Exact Generation: Cholesky decomposition for $O(N^3)$ ground-truth validation.
• Streaming ($O(N^2)$): CachedFGNGenerator for real-time, online noise generation (Incremental Cholesky).

Quantitative Finance

• Rough Processes: fractional_ou_process (Fractional Ornstein-Uhlenbeck) for volatility modeling.
• Asset Pricing: geometric_fbm for simulating asset paths with long memory.
• Constraints: reflected_fbm and fractional_brownian_bridge for boundary-constrained modeling and data imputation.
• Stationarity: fractional_diff (FracDiff) for making financial time series stationary while preserving memory.

Deep Learning & Diffusion

• Noisy Layers: FBMNoisyLinear for replacing standard weights with correlated noise.
• Positional Embeddings: FractionalPositionalEmbedding for Transformers on fractal data.
• Diffusion Tools: SpectralConsistencyLoss to enforce $1/f^\beta$ statistics and HurstScheduler for annealing roughness during sampling.
• Neural SDEs: NeuralFSDE solver with learnable Hurst parameters.

---

Install

From PyPI:

pip install torch-fbm

For Development:

git clone https://github.com/Coder9872/torch-fbm.git
cd torch-fbm
pip install -e .

---

Quick Usage

1. Generate Rough Paths (Batch)

Generate fractional noise on CUDA using the fast Davies-Harte method.

import torch
from torchfbm import fbm

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Generate 4 paths of length 1024 with H=0.7 (Trending/Smooth)
path = fbm(n=1024, H=0.7, size=(4,), method='davies_harte', device=device)

2. Real-Time Streaming (Online)

Use CachedFGNGenerator for tick-by-tick simulation (e.g., Live Trading Environment).

from torchfbm.online import CachedFGNGenerator

stream = CachedFGNGenerator(H=0.3, device=device) # H=0.3 (Rough/Mean Reverting)

for i in range(100):
    val = stream.step() # Returns next point in O(N^2)
    print(f"Tick {i}: {val.item():.4f}")

3. Deep Learning (Regime-Aware Layers)

Replace standard nn.Linear with FBMNoisyLinear.

from torchfbm import FBMNoisyLinear

# Initialize layer with H=0.5 (Standard)
layer = FBMNoisyLinear(32, 10, H=0.5, device=device)

# Dynamic Regime Switching
layer.H = 0.2  # Switch to Rough/Anti-correlated noise
layer.refresh_noise_stream()
y = layer(torch.randn(8, 32, device=device))

4. Generative Diffusion (Hurst Scheduling)

Anneal the roughness of noise during the diffusion reverse process.

from torchfbm.schedulers import get_hurst_schedule

# Start rough (exploration), end smooth (refinement)
hs = get_hurst_schedule(n_steps=1000, start_H=0.3, end_H=0.7, type='cosine')

for t in reversed(range(1000)):
    current_H = hs[t]
    # Use current_H for sampling noise...

5. Financial Processes

Simulate Geometric fBm (Stock Prices) and Fractional OU (Volatility).

from torchfbm import geometric_fbm, fractional_ou_process

# Stock Price Simulation
s = geometric_fbm(n=1000, H=0.7, mu=0.05, sigma=0.2, s0=100.0, device=device)

---

Analysis Tools

from torchfbm import estimate_hurst, fractional_diff

# Differentiable Hurst Estimation
H_est = estimate_hurst(path.unsqueeze(0), min_lag=4, max_lag=64)

# Fractional Differentiation (Stationarity + Memory)
stationary_ts = fractional_diff(path, d=0.4)

---

Notes

• Methods: Use method='davies_harte' for large simulations. Use method='cholesky' for exact validation.
• Stability: $H$ is clamped to $[0.01, 0.99]$.
• License: MIT License.

---