Alpha / Frameworks

Quantitative Research · June 2026

Three advanced mathematical frameworks, weighed honestly against the alpha they actually deliver.

A structured comparison of fractional calculus, topological data analysis, and information geometry for quantitative alpha generation — covering theory, the 2020–2026 academic literature, computational complexity, HFT versus macro feasibility, and a production-ready Python stack. Every empirical claim is cited; speculative claims are flagged.

At a glance

01 / FRACTIONAL CALCULUS

Long memory & rough dynamics

Non-integer derivatives capture the persistent, non-Markovian structure that classical Itô calculus erases.

Key Hurst
H ≈ 0.1
Best fit
Macro vol
Complexity
O(N²) → O(N log N)
Evidence
Strong (pricing)
02 / TOPOLOGICAL DATA ANALYSIS

The shape of the market

Persistent homology detects regime fragmentation and bubble precursors invisible to correlation matrices.

Key signal
L¹‑norm
Best fit
Daily macro
Complexity
O(2ⁿ · k³)
HFT viable
No
03 / INFORMATION GEOMETRY

Distance between regimes

Fisher metric, transfer entropy, and Wasserstein distance turn distribution shifts into measurable signals.

Key tool
Fisher–Rao
Best fit
HFT + macro
Complexity
O(N) to O(N²)
Sharpe lift
+0.15–0.30
01

Fractional calculus & rough volatility

Non-Markovian · Long memory

1.1 — Underlying theory

Fractional calculus generalizes differentiation to non-integer orders. The three workhorse definitions in quantitative finance are the Riemann–Liouville, Caputo (preferred for initial-value problems), and Grünwald–Letnikov derivatives. All are non-local convolution operators: the value of \(D^{\alpha} f(t)\) depends on the full history of \(f\), not just its local behaviour — which is precisely why they matter for markets that remember.

Caputo derivative of order α ∈ (0,1)
\[ {}^{C}\!D_a^{\alpha} f(t) \;=\; \frac{1}{\Gamma(1-\alpha)} \int_a^{t} (t-s)^{-\alpha}\, f'(s)\, ds \]

Hurst exponent & long memory

The Hurst exponent \(H \in (0,1)\) parameterises self-similarity. \(H = 0.5\) is Brownian motion; \(H > 0.5\) is persistent (trending); \(H < 0.5\) is anti-persistent or rough. Empirically, signed order flow exhibits H ≈ 0.7 (strong long memory), while realized volatility is rough with H ≈ 0.1 — the foundational stylized fact behind the rough volatility paradigm (Gatheral, Jaisson, Rosenbaum 2018; Gould, Porter, Howison 2015).

Fractional Brownian motion & rough volatility

Fractional Brownian motion \(B^H\) is the Gaussian process with covariance \(\mathbb{E}[B^H_s B^H_t] = \tfrac{1}{2}(|s|^{2H} + |t|^{2H} - |s-t|^{2H})\). For \(H \ne 0.5\) it is neither Markov nor a semimartingale, requiring Malliavin calculus extensions. Modern rough-vol models drive log-volatility with fBm of small \(H\):

Rough Bergomi log-variance
\[ \log V_t \;=\; \xi_0(t) + \nu \int_{-\infty}^{t} K(t-s)\, dB^H_s, \quad K(t) = \sqrt{2H}\, t^{H-1/2} \]

Why fractional models matter

  • Long memory — power-law autocorrelations replace GARCH/Heston's exponential decay, matching data.
  • Non-Markovian dynamics — future volatility depends on the entire past, not just the current state.
  • Heavy tails — Mittag-Leffler waiting-time CTRWs produce fat tails without ad-hoc tuning.
  • Microstructural foundation — Hawkes-process scaling limits link \(H_\text{flow} \approx 0.75\) to rough volatility \(H_\text{vol} \approx 0.1\) (Muhle-Karbe et al. 2026).

1.2 — Academic literature, 2020–2026

Eighteen papers organised into four tracks. Every citation links to the primary source.

Rough volatility — core

2018
Volatility is Rough

Gatheral, Jaisson, Rosenbaum · Quantitative Finance 18(6)

The foundational empirical paper. Log-volatility behaves as fBm with H ≈ 0.1 at every scale; introduces the RFSV model and proves rough processes can mimic apparent long memory in statistical tests.

FoundationalDOI
2016
Pricing Under Rough Volatility

Bayer, Friz, Gatheral · Quantitative Finance 16(6)

Introduces the rough Bergomi (rBergomi) model — a three-parameter fBm-driven stochastic volatility model with remarkable fit to empirical implied vol surfaces. The reference computational benchmark.

PricingDOI
2019
The Characteristic Function of Rough Heston Models

El Euch, Rosenbaum · Mathematical Finance 29(1)

The classical Riccati ODE becomes a fractional Riccati equation. Enables semi-analytic Fourier pricing under rough Heston and links rough vol to nearly-unstable Hawkes microstructure.

PricingDOI
2024
Statistical Inference for Rough Volatility: Minimax Theory

Chong, Hoffmann, Liu, Rosenbaum, Szymanski · Annals of Applied Probability

Proves the optimal convergence rate for estimating H from n observations is n^(-1/(4H+2)) — substantially slower than parametric rates, quantifying how genuinely hard rough estimation is.

TheoryarXiv
2024
Multivariate Rough Volatility

Dugo, Giorgio, Pigato · arXiv:2412.14353

Extends rough volatility to multivariate fractional Ornstein-Uhlenbeck with asset-specific Hurst exponents and non-trivial cross-covariance. Validated on the full Oxford-Man realized library.

MultivariatearXiv
2023
Detecting Rough Volatility: A Filtering Approach

Damian, Frey · arXiv:2302.12612

Uses fBm's superposition representation as OU processes to develop a filter-based method for identifying roughness directly from observable prices — practical estimation for discrete data.

EstimationarXiv

Fractional SDEs & option pricing

2020
Hierarchical Adaptive Sparse Grids and QMC for rBergomi

Bayer, Ben Hammouda, Tempone · Quantitative Finance

Adaptive sparse-grid quadrature + QMC + Brownian bridge construction for rBergomi Monte Carlo. Substantial speedups over naive simulation without loss of accuracy.

NumericsDOI
2020
Anomalous Diffusions in Option Prices

Jacquier, Torricelli · Finance and Stochastics

CTRWs with Mittag-Leffler waiting times — a fractional Fokker-Planck setting — reproduce the declining implied vol term structure that standard Lévy and SV models miss.

PricingarXiv
2024
Fractional Black–Scholes with Physics-Informed Neural Networks

Yu, Zheng, Song, Luo, Tan · Scientific Reports

PINN for free-boundary problems under Caputo, Caputo–Fabrizio, and Atangana–Baleanu-Caputo derivatives. Yields more realistic American put pricing than classical BS.

ML+FracPMC
2024
Unsupervised Calibration of Rough Bergomi

Teng, Li · arXiv:2412.02135

Unsupervised deep-learning scheme for calibrating rBergomi without labelled training data; jointly learns the BSDE solution and model parameters — bypasses expensive Monte Carlo precomputation.

ML+FracarXiv

Long memory in microstructure

2015
Long Memory of Order Flow in FX Spot Market

Gould, Porter, Howison · arXiv:1504.04354

Clean empirical evidence that signed order flow has H ≈ 0.7 across three FX pairs — estimable within a single trading day, isolating microstructural persistence.

FoundationalarXiv
2026
A Unified Theory of Order Flow, Market Impact, and Volatility

Muhle-Karbe, Ouazzani Chahd, Rosenbaum, Szymanski

Hawkes-process scaling limit ties signed order-flow persistence (H₀), rough volatility (2H₀ − 3/2), and square-root market impact to a single estimable parameter H₀ ≈ 3/4. The most ambitious recent synthesis.

2024
Convergence of Heavy-Tailed Hawkes Processes and Microstructure of Rough Volatility

Horst, Xu, Zhang · arXiv:2312.08784

Weak convergence of nearly-unstable Hawkes processes with power-law kernels to rough Heston. Extends the Hawkes → rough vol bridge to the empirically relevant heavy-tailed case.

TheoryarXiv
2024
Fractional SV Models for Microstructure & Optimal Execution

Webb · Frontiers in Applied Mathematics

Reviews FSV models for execution algorithms; the Hurst parameter interpolates between diffusion and rough regimes with direct implications for transaction-cost estimation.

ExecutionFrontiers

Machine learning + fractional hybrids

2022
Combining Fractional Derivatives and Machine Learning: A Review

Mallinger, Raubitzek, Neubauer · Entropy 25(1)

Systematic taxonomy of fractional+ML approaches: fractional differencing for stationarity, fractional dynamics for physics-informed learning, fractional-order optimization. Maps directly to López de Prado's fracdiff feature engineering.

ReviewDOI
2021
Why Do Big Data and Machine Learning Entail the Fractional Dynamics?

West, Niu, Chen · Entropy 23(3)

First-principles argument that optimal ML over heavy-tailed data requires fractional-order operators; demonstrates how fractional dynamics emerge in stochastic configuration networks under heavy-tailed noise.

TheoryPDF

1.3 — Computational complexity

The fundamental bottleneck is non-locality: every fractional derivative is a convolution over the full history. Unlike standard SDEs (Euler–Maruyama is O(1) per step), fractional schemes are naively O(N²), scaling to O(M·N²) for Monte Carlo with M paths and N steps.

MethodTimeMemoryNotes
Naive GL/RL (full history)O(N²)O(N)Convolution over all prior points
L1 scheme (Caputo)O(N²)O(N)Piecewise-linear quadrature; standard workhorse
Fast L1 / exp-sumO(N log N)O(log N)Kernel as sum of exponentials → auxiliary ODEs
Short-memory truncationO(N·L)O(L)Lose exactness; OK if H not near 1
FFT-based GLO(N log N)O(N)Circulant convolution; exact; uniform grid
Markovian lifting (Abi Jaber)O(kN)O(k)k OU factors approximate rough kernel; restores Markov

Mitigation strategies

  • Hybrid scheme (Bennedsen–Lunde–Pakkanen 2017) — standard for rBergomi MC.
  • Markovian lifting — k = 5–10 OU factors recovers most accuracy at O(N) cost.
  • Neural surrogate — offline-trained network produces millisecond rBergomi calibration in inference.

GPU feasibility

Naive fractional operators are sequential per path — poor GPU fit. But path-level parallelism (Monte Carlo over M paths) is embarrassingly parallel; cuFFT makes FFT-based methods scale near-linearly; neural surrogates are pure GPU forward passes. Direct fractional computation is not real-time feasible at meaningful path counts; surrogates and fractional differencing (offline) are.

1.4 — HFT versus macro

HFT

Mixed — feasible but contested

Microstructure long memory in order flow (H ≈ 0.7) is real and well-documented. Rough intraday volatility forecasts beat GARCH. But estimating H in real time has slow minimax rates, and alpha in liquid instruments is largely competed away.

  • Order-flow persistence signals
  • Intraday rough vol forecasting
  • Markovian-lifted models for risk
  • Caveat: wide confidence intervals; latency demands surrogates

Macro

Strong — primary habitat

Long-horizon vol forecasting, term structure modelling, and pricing of long-dated derivatives are where fractional models have a demonstrated edge over GARCH/Heston. The advantage grows with horizon.

  • 1d–multi-month realized vol forecasts (RFSV)
  • Implied vol term structure (T^(H-1/2) skew)
  • VIX and variance swap modelling
  • Long-dated derivatives, life insurance

1.5 — Alpha generation, honestly assessed

ClaimEvidenceSource
Rough vol fits SPX surface better than standard SV Strong — replicated across data/venues Bayer 2016
RFSV beats HAR/GARCH for realized vol Moderate–strong — major indices Gatheral 2018
Signed order flow long memory (H ≈ 0.7) Very strong — multi-market replication Gould 2015
Fractional differencing preserves ML signal Moderate — backtests, regime-dependent López de Prado 2018
Live trading Sharpe lift from rough-vol strategies Unproven — no peer-reviewed live PnL

Bottom line: fractional calculus delivers real, documented value in options market-making, volatility-product pricing, and realized-volatility forecasting. Its contribution to directional alpha is plausible but largely undemonstrated in transaction-cost-adjusted terms. Most actionable today: rough-vol-informed delta/vega hedging, fractional differencing in ML pipelines, and vol forecasting feeding volatility risk premium strategies.

02

Topological data analysis & market regimes

Persistent homology · Crash precursors

2.1 — Underlying theory

Correlation and PCA collapse data to scalar summaries in Euclidean space; they erase the shape of the joint distribution. TDA preserves global connectivity. Given a point cloud \(X \subset \mathbb{R}^n\), one builds a nested filtration of simplicial complexes parameterised by a scale \(\varepsilon\):

Filtration
\[ \emptyset = K_0 \subseteq K_1 \subseteq \cdots \subseteq K_N = K \]

Simplicial complexes

Vietoris–Rips connects any set of points whose pairwise distances are all ≤ ε; size controlled by clique structure of a proximity graph. Čech requires common intersections of ε-balls; the nerve theorem guarantees correct homotopy but at higher cost. Alpha complexes via Delaunay triangulation are fastest in ≤ 3D.

Persistent homology & Betti numbers

As ε grows, topological features (components, loops, voids) are born and later die. The Betti numbers count features by dimension: \(\beta_0\) = components, \(\beta_1\) = loops, \(\beta_2\) = voids. The persistence diagram Dgm(X) records each feature's birth-death pair (\(b_i, d_i\)). Points far from the diagonal d = b are long-lived (signal); points near it are noise.

Persistence landscapes

Bubenik's landscapes stack the tent functions of each persistence bar into functions \(\lambda_k:\mathbb{R}\to\mathbb{R}_{\ge 0}\). Their \(L^p\)-norms are the workhorse crash signal in the financial literature.

Distances

The bottleneck distance is the optimal-matching \(L^\infty\) cost; Wasserstein distances generalise to \(L^p\). Both are stable: small data perturbations yield bounded diagram changes — essential for noisy markets.

Takens embedding

A univariate return series becomes a point cloud via sliding-window delay embedding \((r_t, r_{t+\tau}, \ldots, r_{t+(d-1)\tau})\). By Takens' theorem this reconstructs a diffeomorphic image of the attractor — cyclic dynamics generate prominent 1-cycles; pre-crash geometry collapses or expands abruptly.

Why TDA catches what PCA misses

PCA assumes ellipsoidal distributions and cannot register non-convex clusters, circular structure, or the connectivity of correlation networks. Persistence diagrams register fragmentation events (new components) and cycle growth that precede crashes — manifesting as \(L^1\)-norm growth ~250 trading days before 2000 and 2008 (Gidea & Katz 2018).

2.2 — Academic literature, 2018–2026

Foundational work

2018
Topological Data Analysis of Financial Time Series: Landscapes of Crashes

Gidea & Katz · Physica A 491

The founding empirical paper. L^p-norms of persistence landscapes show monotonic growth ~250 trading days before the dot-com and Lehman crashes across S&P 500, DJIA, NASDAQ, Russell 2000. Establishes the sliding-window + Rips + landscape pipeline.

FoundationalDOI
2020
Topological Recognition of Critical Transitions in Time Series of Cryptocurrencies

Gidea, Goldsmith, Katz, Roldán, Shmalo · SSRN

TDA + k-means on BTC/ETH/LTC/XRP before the December 2017 crash. Validated first on a Lorenz attractor; early warning signals appear despite high crypto noise. Establishes the TDA+unsupervised-ML pipeline.

CryptoSSRN

Bubble detection & crash precursors

2023
Why Topological Data Analysis Detects Financial Bubbles?

Akingbade, Gidea, Manzi, Nateghi · SSRN

Proves analytically that whenever the LPPLS bubble model fits prices, TDA generates early-warning signals. The first paper to ground TDA crash detection in a theoretical bubble model; validated on Bitcoin's positive and negative bubbles.

TheorySSRN
2020
Time-Resolved TDA of Market Instabilities

Katz & Biem · SSRN

TDA on CDS spreads + equity prices for 93 NA companies. CDS market leads equities in the L¹-norm signal. Critical caveat: the method explicitly fails for exogenous shocks like COVID-19.

CaveatSSRN
2022
A Persistent-Homology-Based Turbulence Index

Ruiz-Ortiz, Gómez-Larrañaga, Rodríguez-Viorato · arXiv:2203.14623

New TDA turbulence index tested across S&P 500, Russell 2000, Mexican IPC, Nikkei 225 through 1987 Black Monday, dot-com, GFC, and COVID-19. Extends validation to non-US markets and earlier crisis epochs.

Multi-marketarXiv

COVID-19 as a test case

2024
Identifying Extreme Events in the Stock Market: A Topological Data Analysis

Rai, Sharma, Luwang, Nurujjaman, Majhi · Chaos 34, 103106

Continent-level + sector-level TDA. L¹, L², and Wasserstein distances all exceed μ+4σ during 2008 and COVID across Americas, Europe, Asia. Indian banking/automobile sectors show prolonged stress; pharma/FMCG do not.

Multi-sectorAIP
2021
Global Stock Markets' Connections with Emphasis on COVID-19

Yu, Guo, Zhang, Zhao · Physica A 580

TDA across 40 global indices, 1995–2020. Market correlations during COVID-19 spread were more extreme than in any prior crisis in sample — including LTCM, dot-com, and the GFC.

ValidationPMC

Cryptocurrency & digital assets

2021
Using TDA & Persistent Homology to Analyze Singapore & Taiwan Markets

Yen & Cheong · Frontiers in Physics 9

Comprehensive demonstration of the full TDA toolkit (barcodes, persistence diagrams, persistent entropy, bottleneck, Betti, Euler) on STI and TAIEX through the 2020 COVID crash. Pedagogically the most complete treatment of TDA diagnostics in one paper.

PedagogyFrontiers
2025
Hierarchical Persistence Velocity for Network Anomaly Detection

Khormali · arXiv:2512.14615

Introduces OW-HNPV — the first velocity-based perspective on persistence diagrams. 10.4% AUC gain over baselines for 7-day Ethereum price-movement prediction, with proven mathematical stability.

SOTAarXiv

Portfolio construction via TDA

2020
Topological Data Analysis in Investment Decisions

Goel, Pasricha, Mehra · Expert Systems with Applications 147

The foundational portfolio paper. Uses Takens + sliding-window persistence to filter assets, then solves an optimization for enhanced indexing. TDA-filtered EI delivers superior excess returns and tail metrics across 10 global datasets.

PortfolioDOI
2025
Sparse Portfolio Selection via TDA-Based Clustering

Goel, Filipović, Pasricha · Quantitative Finance

Sparse index-tracking and Markowitz with TDA clustering and time-aware persistence distances. S&P 500 2009–2022 (incl. COVID) — significant out-of-sample improvement across performance measures, holding up in turbulent markets. The most rigorous TDA portfolio paper to date.

PortfolioDOI
2023
Sparse Index Tracking via Topological Learning

Goel, Pasricha, Kanniainen · arXiv:2310.09578

Vietoris–Rips filtration measures asset "riskiness" and tunes regularization for sparse tracking. Beats Elastic-Net and SLOPE across risk, risk-adjusted, and cost metrics over 23 years of S&P 500 data.

PortfolioarXiv
2019
TDA for Portfolio Management of Cryptocurrencies

Rivera-Castro, Pilyugina, Burnaev · ICDM Workshops

Mapper-based portfolio across 1,500+ cryptocurrencies over 6 years — outperforms classic benchmark methods without TDA domain expertise. Demonstrates practitioner accessibility.

MapperIEEE

TDA + deep learning hybrids

2025
Enhancing Financial Time Series Forecasting Through TDA

de Jesus, Fernández-Navarro, Carbonero-Ruz · Neural Comp. & Applications

TDA features (entropy, amplitude, diagram cardinality) feed N-BEATS. Best mean performance and rank across MAPE/MAE/RMSE on 32 datasets, statistically significant at α = 0.10.

ML+TDASpringer
2024
Applications of TDA in Stock Index Movement Prediction

Huang, Xu, Huang, Chen · arXiv:2411.13881

Compares three point-cloud constructions (Takens, correlation, factor). Correlation-based clouds produce excessive false positives; Takens embedding is more robust. Includes reproducible code pipeline.

MethodarXiv

2.3 — Computational complexity

Worst-case Rips complex is exponential in n; matrix reduction is cubic in simplex count. Practical empirical scaling follows power laws with exponents 4–6 for 4D point clouds. The latency wall is the central HFT constraint.

LibraryOptimizationSpeedup
Ripser (Bauer 2019)Implicit coboundary, clearing lemma10–100× over predecessors
GUDHI Alpha complexDelaunay-based O(n^⌈d/2⌉)Best for ≤ 3D
Ripser++ (Zhang 2020)GPU matrix reductionUp to 30× over Ripser, 2× less memory
Sparse Rips (Chazal)(1+ε)-approximationExponential → log-linear asymptotic

Real-time feasibility

Time scaleWindowPer-window (Ripser, H₁)Feasible?
Millisecond (HFT)Cannot complete in < 1 msNo
Second100 ticks~100 msNo
Minute60 bars~1–5 sMarginal
Hour60 bars~1–5 sYes
Daily50–250 bars~1–10 sYes

2.4 — HFT versus macro

HFT

Not viable

Persistent homology at tick or sub-second latency is impossible — minimum wall time is hundreds of milliseconds. TDA can be a meta-signal for intraday strategy switching at minute-bar or hour-bar scale, but never inside the HFT loop.

  • Intraday regime classification at minute/hour bars
  • Offline topology fingerprinting
  • Risk monitoring, not execution

Macro

Strong — natural habitat

Crash early warning (endogenous bubbles), regime identification, cross-asset structure mapping, and portfolio diversification via topological dissimilarity all have peer-reviewed empirical support at daily and weekly horizons.

  • Crash early warning (L¹-norm growth)
  • LPPLS bubble convergence detection
  • CDS-vs-equity Wasserstein divergence
  • TDA-clustered sparse portfolios
  • Sectoral fragility analysis

2.5 — Alpha generation, honestly assessed

ApplicationBest evidenceQualityCaveat
Crash early warning (endogenous) Gidea & Katz 2018; Akingbade 2023 Moderate 2 crashes; window selection in-sample
Crash detection (exogenous) Katz & Biem 2020 Negative TDA explicitly fails for shocks like COVID-19
Portfolio enhanced indexing Goel et al. 2020, 2024 Strong Walk-forward validation across multiple markets
Index tracking improvement Goel et al. 2023 Strong 23-year out-of-sample window vs. Elastic-Net/SLOPE
Deep-learning feature augmentation de Jesus 2025 Moderate Statistically significant at α=0.10; 32 datasets

Bottom line: the strongest evidence is for portfolio construction, not crash timing. The Goel papers represent the gold standard — proper walk-forward validation, multiple markets, including turbulent periods. Crash early-warning works for endogenous bubble dynamics but fails predictably for exogenous shocks. Watch for look-ahead bias, persistence-diagram instability, and unmodeled transaction costs.

03

Information geometry & market entropy

Fisher metric · Transfer entropy · Wasserstein

3.1 — Underlying theory

Information geometry, systematized by Amari and Nagaoka, treats parametric families of probability distributions as differentiable manifolds. Each point on a statistical manifold \(\mathcal{M}\) is a distribution \(p(x; \theta)\). The canonical Riemannian metric is the Fisher–Rao metric:

Fisher information metric
\[ g_{ij}(\theta) \;=\; \mathbb{E}_{p(\cdot;\theta)}\!\left[\frac{\partial \log p}{\partial \theta^i}\,\frac{\partial \log p}{\partial \theta^j}\right] \]

By the Chentsov–Campbell theorem this is the unique monotone Riemannian metric on statistical manifolds — and yields the tightest Cramér–Rao bound. The geodesic length under this metric gives a proper statistical distance between distributions; rapid market regime shifts are anomalously long geodesic jumps.

Dual connections & α-geometry

Statistical manifolds carry a family of torsion-free affine connections \(\nabla^{(\alpha)}\) parameterised by \(\alpha \in \mathbb{R}\). The \(\alpha = \pm 1\) (exponential, mixture) connections are dual w.r.t. the Fisher metric. Natural gradient descent (Amari 1998) preconditions gradients by \(G^{-1}(\theta)\) — the steepest descent in the intrinsic sense.

Divergences: KL, Jensen–Shannon, Bregman

\(D_\text{KL}(p\|q)\) is the archetypal Bregman divergence generated by the negative entropy potential; equals the Fisher metric to second order. Jensen–Shannon is symmetric and bounded; its square root is a metric on the space of distributions. Bregman divergences \(D_\phi(p,q)=\phi(p)-\phi(q)-\nabla\phi(q)^\top(p-q)\) generalise KL and form the algebraic backbone of information geometry.

Distinction from classical entropy

Shannon, Rényi, and Tsallis entropies are scalars applied to empirical distributions. The Fisher metric is a tensor — a k×k matrix capturing local curvature. Classical entropies have no notion of distance between distinct regimes; Fisher–Rao does.

Wasserstein geometry: related but distinct

\(W_2(p,q)^2 = \inf_\gamma \int \|x-y\|^2 d\gamma\). Unlike KL, Wasserstein does not require shared support and is sensitive to location shifts. It lives in the space of distributions over \(\mathbb{R}^n\), not parameter space \(\theta\). Sinkhorn regularization bridges the two by adding a KL penalty. Wasserstein is the workhorse of distributionally robust portfolio optimization.

Why this captures market structure

  • Distance between regimes — geodesic length quantifies "how different are these regimes."
  • Diversification as geometry — diversified portfolios sit far from concentrated ones on the manifold.
  • Natural gradient optimization — Fisher-preconditioned updates converge faster for distribution fitting.
  • Model comparison — rolling KL/JS divergence detects structural breaks without a breakpoint hypothesis.

3.2 — Academic literature, 2020–2026

Information geometry of volatility families

2025
An Approach to Fisher-Rao Metric for Infinite-Dimensional Non-Parametric Information Geometry

Cheng & Tong · Entropy 28(4)

Resolves infinite-dimensional Fisher–Rao intractability via Structural Decomposition of the Tangent Space (covariate FIM). Trace Theorem H_G(f) = Tr(G_f) establishes G-entropy as a geometric invariant — applicable to non-parametric GARCH manifolds.

TheoryDOI
2021
Information-Theoretic Measures and Modeling Stock Market Volatility

Nasir & Sheraz · Risks 9(5)

Compares EGARCH, GJR-GARCH, TGARCH with Shannon, Tsallis, approximate, and sample entropy on Pakistan Stock Exchange. Pandemic-period volatility manifests as cross-entropy spikes — entropy provides complementary info to GARCH parameters.

EmpiricalPDF

Transfer entropy in finance

2024
Inferring Dealer Networks in FX Using Conditional Transfer Entropy

Janczewski, Kandhai, Anagnostou · Entropy 26(9)

Conditional TE on millisecond bid/ask data reconstructs FX dealer information networks around ECB announcements. Identifies dealers as sources vs. conduits; shows information asymmetry during stress events.

HFTDOI
2020
The Flow of Information in Trading: An Entropy Approach to Market Regimes

Hawkes, Liu, Chen, Yang · Entropy 22(9)

Conditional block + transfer entropy on 11 years of news and market data identifies return-driven, news-driven, and mixed regimes. Entropy-based classification outperforms threshold methods — one of the most actionable macro regime papers.

MacroPDF
2024
Early Warning of Systemic Risk in Commodity Markets via Transfer Entropy Networks

Wei, An, Gao, Sun, Zhao · Entropy 26(7)

Rolling TE networks on 25 Chinese commodity prices; topology changes (edge density, hub concentration) lead systemic stress events by 2–4 weeks.

Early warningPMC
2021
Information-Theoretic Causality Between Financial and Sentiment Data

Cerchiello, Scaramozzino, Aste · Entropy 23(5)

Transfer entropy on sentiment vs. S&P 500 top-50 (Nov 2018–Nov 2020). Significant asymmetric flows from sentiment to price at daily frequency; strongest during COVID-19 onset. Sharpe lift modest but significant.

SentimentPMC

Permutation, sample, and Tsallis entropy

2022
Regularity in Stock Market Indices within Turbulence Periods: The Sample Entropy Approach

Majewska & Olbryś · Entropy 24(7)

Tests and confirms across 36 European/US indices: sample entropy decreases during 2007–2009 GFC and 2020–2021 COVID. Crisis periods exhibit higher regularity/predictability — a remarkably robust finding.

RobustPDF
2025
Entropy-Assisted Quality Pattern Identification in Finance

Singh, Gupta, Kais, Gupta · Entropy 27(4)

Local permutation entropy prunes low-quality short-term price patterns; tested on Gold/USD and GBP/USD. Entropy-filtered patterns outperform unfiltered in backtests.

HFT-adjPMC
2022
Tsallis Relative Entropy as a Risk Measure for Financial Portfolios

Devi & Page · arXiv:2205.13625

Constructs risk-optimal portfolios using Tsallis relative entropy from asymmetric heavy-tailed return distributions. Over >10-year horizons, TRE portfolios beat CAPM-beta on risk-adjusted returns.

PortfolioarXiv

Wasserstein distance & distributionally robust optimization

2024
Robustifying Conditional Portfolio Decisions via Optimal Transport

Nguyen, Zhang, Wang, Blanchet, Delage, Ye · arXiv:2103.16451

DR mean-variance and mean-CVaR portfolios with Wasserstein ambiguity sets conditioned on covariates. Finite-dimensional reformulations; documented out-of-sample improvement over non-robust baselines.

PortfolioarXiv
2025
Wasserstein Robust Market Making via Entropy Regularization

Fang & Israel · arXiv:2503.04072

Robust market-making under Wasserstein uncertainty as convex optimization. Combines KL entropy regularization with Wasserstein robustness in one framework; data-driven radius selection. Directly applicable to HFT quote optimization.

Information-theoretic causal discovery

2020
Synergistic Information Transfer in the Global System of Financial Markets

Marinazzo, Faes, Mantegna, Stramaglia, Scagliarini · Entropy 22(9)

Introduces synergy — a higher-order information measure beyond bivariate Granger causality — to characterise three-variable circuits. European-American index pairs show the strongest synergetic circuits.

TheoryPMC
2024
On the Three Demons in Causality in Finance: Time Resolution, Nonstationarity, Latent Factors

Dong, Dai, Fan, Jin, Rajendran, Zhang · arXiv:2401.05414

Systematically addresses three TE failure modes in financial time series and proposes solutions for each. Essential reading for practitioners implementing transfer-entropy pipelines.

MethodarXiv

3.3 — Computational complexity

MethodComplexityNotes
Shannon/Rényi entropy (histogram)O(N)Bin choice critical; fast
Permutation entropy (order m)O(N log N)Sort-dominated; real-time viable
Approximate entropy (ApEn)O(N²)Template matching; N ≤ 5000 practical
Sample entropy (SampEn)O(N²)kd-tree → O(N log N) in practice
Transfer entropy (binned)O(N log N)Binning + histogram
TE via kNN (Kraskov)O(N²)More accurate; bottleneck N > 10⁴
Fisher metric (parametric)O(k²N + k³)Fit + invert; k = param dim, small
KL divergence (empirical)O(N log N)kNN or histogram; well-studied
Wasserstein exact (LP)O(N³ log N)Impractical N > 500; avoid
Sinkhorn approximationO(N² / ε)ε = regularization; GPU-friendly

For univariate tick-level data (N ~ 10³–10⁴), permutation entropy, sample entropy, and Shannon TE are all microsecond-to-millisecond — viable for HFT order-flow analysis. Multivariate TE networks are O(d²N²); practical cap ~50 assets for daily recomputation without GPU. Wasserstein at N = 252 with Sinkhorn takes ~10 ms on CPU — fine for daily rebalancing.

3.4 — HFT versus macro

HFT

Genuinely viable

Permutation entropy on order-flow imbalance, transfer entropy for sub-second dealer lead-lag, depth entropy for liquidity forecasting, and Wasserstein-robust market making are all computationally tractable at HFT scales — a clear advantage over TDA and fractional calculus.

  • Order-flow PE for adverse selection
  • Sub-second TE lead-lag networks
  • Depth entropy > bid-ask spread proxy
  • Wasserstein-robust quote optimization

Macro

Strong

Sample entropy regime detection, cross-asset TE contagion networks, KL/Wasserstein structural-break tests, and entropy-portfolio diversification all have replicated empirical support.

  • Entropy spikes for regime detection
  • TE contagion networks (commodity, equity, FX)
  • Rolling KL for distribution shifts
  • Effective Number of Bets = exp(H(w))
  • Wasserstein DRO portfolios

3.5 — Alpha generation, honestly assessed

SignalHorizonEvidenceIndicative Sharpe lift
Sample/PE as regime filterDaily–weekly Strong — replicated across crises +0.15 to +0.30 vs. buy-and-hold
TE for sentiment→priceDaily Moderate — Cerchiello 2021 Modest; signal decays post-COVID
TE commodity early warningWeekly Moderate — Wei 2024 Sharpe ~0.6–0.8 vs. ~0.4 baseline
Entropy-filtered HFT patternsSub-daily Emerging — Singh 2025 Framework shown; live PnL not published
Wasserstein DRO vs. 1/NMonthly Moderate — Zhou & Liu 2024 Better CVaR; Sharpe lift marginal
Geodesic anomaly detectionDaily Speculative — no peer-reviewed PnL

Where signals are likely priced in

Information theory does not exempt markets from efficiency. Stable, publicly-known TE signals should be arbitraged. The documented signals tend to be short-lived (regime-specific) or require data that isn't easily accessible (dealer-level FX quotes, granular order flow). Dong et al. (2024) formalise how non-stationarity and latent factors corrupt TE estimates — a TE network estimated on 2019 data may be qualitatively incorrect in 2022.

Bottom line: information geometry covers the broadest set of computationally viable alpha applications among the three frameworks — but most published signals are gross of transaction costs, and pure Fisher-manifold geodesic signals remain theoretically elegant but empirically unvalidated. Stick to the well-supported quadrants: entropy-based regime filters, transfer-entropy networks, Wasserstein-robust market making, entropy-portfolio diversification.

Side-by-side structured comparison

One matrix, every dimension that matters for deployment decisions.

Dimension Fractional Calculus Topological Data Analysis Information Geometry
Core mathematical object Non-integer-order convolution operator Persistent homology of filtered simplicial complex Riemannian manifold of probability distributions
Key signal Hurst exponent H; fBm-driven volatility L¹/L²-norm of persistence landscape; Wasserstein on diagrams Fisher metric; KL/JS/Wasserstein divergence; transfer entropy
What it captures Long memory, non-Markovian dynamics, heavy tails Shape, fragmentation, cycles, regime transitions Distributional distance, causal flow, complexity
Best Big-O O(N log N) via FFT or exp-sum O(2ⁿ · k³) Rips; O(n^⌈d/2⌉) Alpha for d ≤ 3 O(N log N) for most univariate measures
HFT latency feasibility Surrogates only (neural / Markovian lifting) Not viable — minute-bar minimum Yes — PE/TE viable at sub-second
Macro feasibility Strong — primary habitat Strong — primary habitat Strong — broad applicability
GPU acceleration Path-parallel MC, cuFFT, neural surrogates Ripser++ (up to 30×); limited otherwise geomloss / Sinkhorn → strong; PE/SampEn CPU-bound
Strongest evidence SPX implied vol surface fit; realized vol forecasting Sparse portfolio construction (Goel 2020–2025) Sample entropy regime detection; TE early warning
Known failure mode Slow H estimation (n^(-1/(4H+2))); not real-time Exogenous shocks (COVID); look-ahead in window choice Non-stationarity corruption; pricing-in of public signals
Pre-2020 maturity Mature (rough vol since 2014) Emerging (Gidea–Katz 2018 foundational) Mature in theory; applied finance ramp 2018+
Production track record Yes — quant desks use rough Bergomi for SPX/VIX Limited — mostly academic; some hedge fund use Yes — entropy regime filters in risk parity products
Honest alpha verdict Pricing/hedging edge real; directional alpha unproven Portfolio construction edge strong; crash timing nuanced Broadest set of viable signals; most published gross-of-cost
Top library fracdiff, fbm giotto-tda, ripser POT, antropy

Python implementation stack

Production-relevant libraries organised by framework. Versions and notes reflect the 2024–2026 ecosystem.

Fractional calculus

Long memory · Rough volatility
López de Prado-style fractional differencing via FFT convolution. sklearn-compatible Fracdiff / FracdiffStat (auto-selects minimum d for ADF stationarity). PyTorch module for differentiable pipelines. ~10,000× faster than reference implementations.
pip install fracdiff
Pure-Python numerical fractional calculus. Implements Grünwald–Letnikov, Riemann–Liouville, and three Caputo variants (L1, L2, L2C). Naive O(N²) — suitable for research and moderate series.
pip install differint
Exact fBm and fractional Gaussian noise simulation via Hosking, Cholesky, and Davies–Harte (fastest, O(N log N) FFT-based). The go-to library for rough volatility Monte Carlo research.
pip install fbm
Broader process library: fBm, GBM, Ornstein–Uhlenbeck, Bessel, Cauchy, and more. Euler–Maruyama for diffusions; exact for fBm. Good for multi-asset stochastic prototyping.
pip install stochastic
Lightweight R/S analysis for Hurst exponent estimation. Returns H, constant c, and regression data for diagnostic plots. Needs ≥100 points for reliable estimation.
pip install hurst
Oxford DataSig group's path-signature library — universal features for sequential data, grounded in rough path theory (Lyons). C++ backend; ideal for path-dependent payoff approximation and signature regression in ML.
pip install RoughPy
Time-series fractional differencing with auto-selection of optimal d. scikit-learn pipeline compatible. Useful complement to fracdiff for batch feature engineering.
pip install tsfracdiff

Topological data analysis

Persistent homology · Regimes
Most comprehensive TDA + ML library. Integrates persistent homology and Mapper into sklearn Pipelines. C++ backends (GUDHI, Ripser, flagser). Handles tabular, time series, graphs, and images natively. Top pick for financial TDA.
pip install giotto-tda
Python wrapper around Bauer's Ripser. Best-in-class speed for Vietoris–Rips barcodes in dim ≥ 2. Supports sparse inputs, lower-star filtrations on images, and representative cochains.
pip install ripser
Inria's reference TDA library. Rips, Čech, alpha complexes; persistence cohomology; bottleneck distance; sklearn-compatible interface (v3.10+, 2024 NanoBind rebuild). Alpha complex fastest for ≤3D data.
pip install gudhi
Persistence diagram comparison and vectorization: bottleneck distance, sliced Wasserstein kernel, heat kernel, persistence images. Essential downstream tool.
pip install persim
Flexible Mapper algorithm with D3.js / Plotly visualizations. Leverages scikit-learn for lens and clustering. Standard choice for Mapper-based portfolio construction.
pip install kmapper
Deep-learning extension for giotto-tda. Topological loss functions and PersLay-style layers for PyTorch. Enables end-to-end differentiable topological-feature pipelines.
pip install giotto-deep
Ripser++ (GPU)
GPU-accelerated Vietoris–Rips computation (Zhang 2020). Up to 30× speedup, 2× memory reduction. Research release; not yet on PyPI as a stable package.
github · research release

Information geometry

Fisher · Entropy · Wasserstein
Vallat's Numba-JIT entropy toolkit. Permutation, sample, approximate, spectral, SVD, Lempel–Ziv entropies; Hjorth parameters. Fast, BSD, actively maintained. Top pick for HFT-grade entropy.
pip install antropy
Most comprehensive entropy toolkit (v2.0, 2024). 20+ base entropies (ApEn, SampEn, FuzzEn, PermEn, DispEn, etc.), cross-entropies, multiscale, bidimensional, multivariate. WindowData() for rolling computation.
pip install EntropyHub
The reference Wasserstein library. Exact OT (LP), Sinkhorn (entropic, log-domain), barycenters, Gromov–Wasserstein, unbalanced OT, sliced OT. PyTorch/JAX/TF backends. Top pick for Wasserstein-DRO portfolios.
pip install POT
GPU-accelerated Sinkhorn divergences via KeOps. Scales to N ~ 10⁶ and differentiable through PyTorch. Essential for large-scale online Wasserstein comparisons in HFT.
pip install geomloss
Statistical manifolds, SPD matrices, Lie groups, information manifold with Fisher metric, geodesics, natural gradient. sklearn API; NumPy/PyTorch/Autograd backends. Top pick for Fisher-Rao geometry.
pip install geomstats
Schreiber's original transfer entropy, active information storage, mutual information, block entropy. Python with C core. Fast for discrete signals — standard for TE network construction.
pip install pyinform
Riemannian optimization in PyTorch: SGD and Adam on manifolds. Natural gradient as a first-class optimization primitive. Pairs with geomstats for end-to-end Fisher-geometric portfolio optimization.
pip install geoopt

Recommended starter stacks

For HFT / tick-level

# Microstructure entropy + path signatures
pip install antropy        # PE on order flow (sub-ms via Numba)
pip install pyinform       # Sub-second TE lead-lag networks
pip install geomloss       # Online GPU Sinkhorn
pip install RoughPy        # Path signatures from tick data
pip install fracdiff       # Offline feature engineering

For macro / multi-asset

# Regime detection + portfolio construction
pip install giotto-tda     # TDA-clustered sparse portfolios
pip install fbm            # Rough vol Monte Carlo
pip install POT            # Wasserstein DRO
pip install EntropyHub     # Rolling sample-entropy regimes
pip install geomstats      # Fisher-Rao on GARCH manifolds
pip install pyinform       # Cross-asset TE contagion networks