Perceptron Intelligence – Research & Collaboration Hub

Comprehensive analysis of Coppock-based technical signals, including: (1) multi-timeframe Coppock strategy design, (2) security-level Coppock scores using voting logic, (3) performance comparison across multiple Coppock variants, and (4) risk/return profiling of the final strategy.

Research comparing rule-based and machine learning approaches for sector trend prediction. Evaluates multiple ML models on technical signals and benchmarks them against a non-ML baseline: a vote-based trade signal using Coppock scores.

Integrated framework for macro regime detection and sector rotation. Combines: (1) ML-based market regime identification, (2) NLP-driven sentiment analysis of news/social media, and (3) unsupervised clustering (k-means, t-SNE) of sector returns to reveal cyclical/defensive, growth/value, and commodity/service dynamics. Builds on Q3 2024 work with heatmap visualizations and cumulative trend analysis for tactical allocation.

Foundational research on macro regime detection using HMM and SHAP-interpretable ML models. Focuses on predicting ETF (sector proxy) returns under varying economic conditions. Integrates NLP for financial text and sentiment analysis, with SHAP values used to explain feature contributions—grounded in cooperative game theory (Shapley, 1953)—to enhance model transparency and strategic decision-making in asset management.

Extension of the vote-count portfolio framework using dynamic signal weighting. Evaluates multiple time series transformations—including Monthly Difference, Rolling Mean (various windows), Exponentially Weighted Moving Average (EWMA), and Expanding Mean—to denoise signals, capture momentum, and adapt vote weights over time for improved portfolio responsiveness and risk-adjusted returns.

Introduces a weighted PCA methodology to derive orthogonal risk factors and forecast asset returns. Portfolios are ranked by weighted PCA scores and grouped into deciles (0–9), with Decile 0 delivering the highest cumulative returns. The approach aligns with Arbitrage Pricing Theory (APT), treating principal components as priced risk factors. Investors implicitly bet on the persistence of associated risk premiums through this construction.

Deep dive into the Vote Count methodology across 500+ factor combinations (span, transformation, cutoff). Monthly portfolios are evaluated for excess returns, information ratio, and hit rate. Includes correlation analysis of raw factor values to identify and prune redundant signals—enabling a leaner, more robust set of inputs for portfolio construction while preserving predictive diversity.

Early integration of Vote Count methodology with ML-driven factor analysis. Uses rolling-window correlations (e.g., yield curve vs. factors) for regime detection, multivariate sequence prediction over 5-year windows, and ML-based feature importance to score factors. Portfolios are formed by vote thresholds (e.g., ≥3 votes), with backtests showing strong outperformance and high information ratios for high-vote portfolios versus low-vote counterparts.

Analysis of factor return distributions with focus on kurtosis as a measure of tail risk. Key findings: (1) normalized momentum factors align with BARRA distribution profiles, (2) growth factors exhibit elevated kurtosis, and (3) kurtosis diminishes when using moving averages longer than 24 months. Includes active return reporting for quantile and cutoff portfolios using winsorization protocols established in February 2023.

Foundational study comparing quantile-based vs. absolute cutoff-based portfolio construction. Evaluates impact of kurtosis, skew, and average factor exposure on returns across overlapping and non-overlapping buckets. Determines optimal tile size (decile vs. quintile) and validates cutoff ranges for robust cross-sectional ranking—establishing core methodology later used in Vote Count and factor scoring frameworks.

Extension of Q4 2022 cutoff methodology using moving average-transformed factor exposures. Compares two winsorization approaches (fixed ±3 vs. percentile-based renormalization) and evaluates octile portfolios via information ratio against benchmark. Establishes a robust, long-only ranking pipeline for cross-sectional factor investing.

Foundational analysis comparing quantile and cutoff portfolios under various moving average transformations of factor exposures. Evaluates performance across tenure-based MA windows to identify robust signal preprocessing methods for long-only cross-sectional strategies. Establishes the baseline methodology later extended in 2023 Q1–Q2 research.

Evaluation of dynamic long-only portfolios using transformed BARRA fundamental factor exposures. Applies moving averages (3M–60M) and moving differences to raw factor data, followed by monthly rebalancing based on transformed scores. Analyzed in xBSL space to assess responsiveness to market regime shifts.

Development of a dual-path factor scoring framework: (1) risk premium approach using cross-sectional regressions of BARRA exposures vs. excess BSL returns, smoothed via moving averages; and (2) direct moving average transformation of factor exposures with grid-search optimization of lookback windows. Evaluated over April–August 2022, forming the basis for dynamic portfolio construction.

Hybrid methodology combining PCA-based dimensionality reduction on tagged BARRA factor subgroups (top 3 eigencomponents per group) with monthly cross-sectional regressions against 3-month forward excess BSL returns to estimate dynamic risk premiums. Derived eigen-variables serve as inputs for portfolio scoring and construction, establishing an early framework for factor synthesis and signal generation.