The Architecture of Compounding: A Physics-Informed, Data-Centric Framework for Geometric Growth

We model compounding not as an incidental outcome but as an engineered architecture: a dynamic system that shapes time, uncertainty, and decision rights to maximize geometric growth. Drawing from non-equilibrium physics, information theory, and CFA Level III portfolio construction, we synthesize a practical framework for institutional capital deployment under real-world frictions.

1) Abstract

Compounding at institutional scale is constrained by drawdowns, liquidity, and non-stationarity. We present a framework in which geometric growth is the objective function, uncertainty is addressed via physics-inspired state estimation, and portfolio decisions are encoded as control operations on a stochastic system. The architecture integrates (i) time-average growth optimization (ergodicity considerations), (ii) information-theoretic budgeting of scarce, costly signal, (iii) fractional Kelly risk governance informed by drawdown tolerances and parameter uncertainty, and (iv) DBA-grade data engineering to prevent leakage and ensure repeatability. We argue this architecture is robust under realistic frictions (impact, latency, constraints) and aligns with the fiduciary mandate to grow wealth at acceptable path risk.

2) Introduction: From Arithmetic to Architecture

Arithmetic returns fetishize point estimates; architecture privileges the distribution of paths. For a continuously rebalanced strategy with log-returns \( r_t \), the time-average growth rate \( g \) approximates \( \mathbb{E}[r_t] - \tfrac{1}{2}\mathrm{Var}(r_t) \) under mild regularity. This decomposition makes explicit the central trade-off: every unit of variance subtracts from growth. Portfolio design therefore becomes an engineering problem: to partition risk and information so that expected return per unit variance is maximized in time, not in a static ensemble snapshot.

3) The Geometry of Growth: Time vs. Ensemble

Classical mean-variance optimization evaluates ensemble moments; investors live one path. In non-ergodic systems, the ensemble average need not equal the time average. If wealth follows multiplicative dynamics \( W_{t+1}=W_t(1+R_{t+1}) \), then maximizing \( \mathbb{E}[\ln(1+R)] \) is the natural objective. This is intimately related to the Kelly criterion in both discrete and continuous time. Importantly, the Kelly solution is path-optimal but also variance-hungry; even small estimation error can translate into intolerable drawdowns. The architectural remedy is fractional Kelly within a risk-budgeted ensemble of edges.

Design implication: treat growth as a geometric objective, then back-solve position sizes and diversification to maximize time-average growth under drawdown constraints.

4) Signal, Entropy, and the Price of Information

Alpha is scarce and costly. In an information-theoretic framing, the mutual information \( I(S;R) \) between a signal \( S \) and future returns \( R \) bounds the extractable Sharpe. For jointly Gaussian variables, \( I(S;R)= -\tfrac{1}{2}\ln(1-\rho^2) \), where \( \rho \) is the correlation between the forecast and realized returns. This suggests two levers: raise \( \rho \) via better features/modeling; or raise breadth (number of independent bets) to translate information into compounding. CFA-level intuition connects here: Information Ratio \( \mathrm{IR} \) scales with skill and breadth; physics analogies interpret increasing breadth as increasing the effective dimensionality of the system, reducing variance per dimension via diversification.

Information is not free: it incurs data engineering cost, capacity cost (alpha decays with scale), and governance cost (more complex models require heavier controls). The architecture therefore prices information by its marginal contribution to geometric growth after costs and constraints.

5) From Kelly to Fractional Kelly: Risk and Drawdown Governance

For a single edge with expected excess return \( \mu \) and variance \( \sigma^2 \), continuous-time Kelly suggests \( f^* \approx \mu/\sigma^2 \). In practice, Sect Capital implements fractional Kelly, \( f = \kappa \cdot f^* \), \( 0<\kappa<1 \), to immunize against parameter error, non-stationarity, and tail risk. Across multiple edges, the Kelly vector is \( \mathbf{w}^* \propto \Sigma^{-1} \boldsymbol{\mu} \), which is formally identical to mean-variance but here interpreted through the growth-optimal lens (weights scale with precision \( \Sigma^{-1} \)).

Governance overlays constrain realized path risk: max drawdown limits, tail-loss limits (e.g., 99% Expected Shortfall), and liquidity-aware circuit breakers. These constraints redefine the feasible Kelly fraction \( \kappa \) dynamically through time, coupling portfolio risk with market microstructure and funding conditions.

6) Control Theory, State Estimation, and Regime Awareness

We cast market states as latent variables evolving under non-stationary dynamics. A Kalman-style or particle filtering approach estimates hidden drivers (e.g., trend strength, volatility regime). The portfolio is a controller that maps state estimates to actions (allocations). Physics contributes both the intuition (non-equilibrium, dissipation) and the tools (state-space modeling, uncertainty propagation). Practically, regime awareness rescales position sizes, adjusts holding periods, and conditions execution style (passive vs. aggressive) to minimize entropy production—i.e., unnecessary turnover and impact.

7) Portfolio Construction: Budgeting Risk for Geometric Objectives

7.1 Risk budgets and edge accounting

We decompose total risk into edge-level contributions. Each edge earns a risk budget proportional to its expected contribution to time-average growth. Budgets are not static; they update as estimated Sharpe, correlation, and crowding evolve.

7.2 Diversification as variance throttle

At a system level, diversification linearizes the growth-variance trade-off: correlated edges are discounted; orthogonal edges are promoted. This controls the \( \tfrac{1}{2}\mathrm{Var}(r_t) \) penalty without sacrificing drift \( \mathbb{E}[r_t] \).

7.3 Constraints as design choices

Leverage caps, concentration limits, and liquidity floors are not mere compliance features; they define the manifold on which compounding occurs. Optimization must be solved on that manifold (e.g., second-order cone formulations for ES constraints).

8) Execution, Capacity, and the Square-Root Law

Implementation transforms paper edges into realized P&L. Empirically, impact often scales with the square root of participation: \( \mathrm{Impact} \approx \sigma \cdot \phi \sqrt{q/V} \) (where \( q \) is trade size and \( V \) market volume). Capacity is found where marginal growth from scaling equals marginal impact and slippage costs. This yields geometric capacity, not just return capacity: the largest size at which time-average growth remains maximal after costs and drawdown risk.

9) Data Architecture: Lineage, Feature Stores, and DBA Discipline

DBA rigor is a first-class growth driver. We standardize:

• Lineage & immutability: raw market data are append-only; transformations are versioned DAGs with checksums; every model artifact is traceable to a data release.
• Event-time vs. processing-time: all features are computed in event-time to avoid look-ahead; late data and revisions handled via slowly changing dimensions (Type 2).
• Feature store: a centralized, permissioned registry with schema contracts (units, timezones, holidays), statistical drift monitors, and deprecation policy.
• Validation harness: unit tests for features, backtests, and risk reports; CI pipelines block promotion on test failure.
• Access controls & audit: principle of least privilege; cryptographic audit trails; reproducible containers for research.

This discipline prevents leakage, supports reproducibility, and dramatically increases the signal-to-noise ratio of research throughput—key inputs to compounding.

10) Validation Protocol: Walk-Forward, Leakage, and Deflated Sharpe

We enforce nested cross-validation with walk-forward evaluation, separating hyperparameter selection from performance estimation. Multiple hypothesis risk is addressed with methods such as the deflated Sharpe ratio (which adjusts the apparent Sharpe for selection bias and non-Gaussianity). White’s reality check and block bootstrap techniques further stress robustness. Out-of-sample degradation is expected; strategies that degrade gracefully under realistic transaction cost models are privileged.

11) Risk Governance: Tail Models, Liquidity, and Fat-Tail Ethics

Tail risks are modeled using a combination of EVT (peaks-over-threshold), conditional ES, and liquidity-aware VaR; stress libraries include historical episodes (e.g., funding squeezes, limit-down regimes) and synthetic shocks (volatility spikes with endogenous liquidity withdrawal). Governance commits to predictable de-risking rules and scenario transparency with LPs, recognizing that long-horizon compounding is a social contract as much as a mathematical one.

12) Sect Capital Principles: Operating the Architecture

• Time-average primacy: optimize \( \mathbb{E}[\ln(1+R)] \), not headline CAGR alone.
• Information is scarce: price signal by mutual information and capacity; scale breadth before leverage.
• Fractional Kelly by design: govern \( \kappa \) via drawdown tolerance, parameter uncertainty, and regime diagnostics.
• Physics for state estimation: non-equilibrium, not static equilibrium; state-space models beat static betas.
• DBA discipline: data lineage and feature governance are compounding levers, not back-office detail.
• Executable research: impact-aware sizing and capacity caps are part of the research spec, not afterthoughts.

13) Conclusion

Compounding is a property of paths under uncertainty. Treating it as an architectural problem—integrating time-average objectives, information-theoretic budgeting, control-theoretic state estimation, and DBA-grade data discipline—yields a portfolio that is measurably more robust. The result is not maximal leverage, but maximal resilience, which is the substrate of geometric growth.

Related Reading

For complementary perspectives, see:
• Markets as Language — narrative, volatility, and syntax of price.
• The Mathematics of Discipline — governance of path risk and repetition.

Disclosures: This content is for information only, not investment advice. Past performance is not indicative of future results. Methods described herein are subject to change without notice.