Research

Methodology & publications

Our strategies are grounded in rigorous quantitative research. Transparency and reproducibility are core principles.

Forthcoming Paper

Multiple Appearance Days (MAD) as an Alpha Indicator in Quantitative Portfolio Construction

George Dikos · Goldflower Capital · 2026

This paper introduces the MAD metric — the number of times an asset appears among top-ranked holdings across multiple independent strategy configurations — as a predictive signal for future returns. Using point-in-time data from 2023 to 2026 (758 trading days), we demonstrate that MAD-based portfolio construction achieves superior risk-adjusted performance compared to single-strategy approaches and the S&P 500 benchmark.

Target: arXiv / SSRN Data: 2023–2026

Paper Sections

1

Introduction

Motivation for the MAD metric and its relationship to consensus-based alpha generation across diverse strategy configurations.

2

Methodology

Streak returns, forward returns, point-in-time SQL, and portfolio construction with T+1 execution. Mathematical framework for rank-based weighting.

3

Results

Backtest tables, top 20 assets by return, performance comparison across weighting schemes, and comparison to S&P 500.

4

Discussion

Statistical significance of MAD as an alpha indicator. Correlation analysis with future returns and comparison versus random selection.

5

Conclusion

MAD as a robust, interpretable signal for systematic portfolio construction with practical implementation guidelines.

Methodology

How the engine works

A transparent breakdown of our portfolio construction pipeline.

1. Strategy Universe Generation

We enumerate strategy configurations across multiple dimensions: lookback windows (5 to 100 days), holding periods (1 to 20 days), portfolio sizes (5 to 30 assets), and benchmark indices (SPY, XITK). Each configuration produces a daily list of top-ranked assets based on historical performance patterns stored in our PostgreSQL database.

2. Winning Strategy Selection (Point-in-Time)

At each evaluation date T, strategies are ranked by their cumulative performance up to and including T — never using future information. We use point-in-time SQL queries against the long_winner_lastday_events table to ensure strict temporal integrity.

Baseline Method: Top 3 by streak return + Top 3 by avg forward 5-day return
Total: 6 winning strategies selected per evaluation date

3. Asset Pooling & Rank-Based Weighting

Holdings from all winning strategies are pooled. Each asset receives a weight inversely proportional to its best rank across all strategies in which it appears.

w(i) = (1 / rank_i) / Σ(1 / rank_j)    for all assets j in pool

Rank 1 (best) → highest weight
Assets appearing in multiple strategies use their best (lowest) rank

4. T+1 Execution Rule

All portfolio signals are executed at the next trading day’s close price. This strict rule eliminates lookahead bias and ensures all backtested returns reflect achievable real-world performance.

Signal generated at close of day T → Trade executed at close of day T+1
Return measured from T+1 close to holding period end

5. Performance Measurement

We report cumulative return, annualized return, annualized volatility, Sharpe ratio (with risk-free rate of 4.75%, the US 10Y average 2023–2026), maximum drawdown, and annualized alpha versus the S&P 500.

Machine Learning

Hyperperformer prediction

Beyond rule-based strategies, we apply ML to identify assets with exceptional return potential.

The ML Pipeline

Our hyperperformer detection system trains on historical asset features — including MAD scores, momentum, volatility, and sector characteristics — to predict which holdings will deliver outsized returns. The model produces a probability score for each asset, and high-confidence predictions are flagged for enhanced allocation.

Key outputs include precision-recall analysis at multiple probability thresholds, allowing us to calibrate the confidence level at which we act on ML predictions.

Optimal Threshold
~0.7
Approximately 80% precision at this level

Research Areas

Time-Split Validation

Train on historical data, predict on out-of-sample forward periods. Ensures the model generalizes to unseen market conditions.

Kelly Criterion Sizing

Uses win probability estimates from the ML model to compute optimal Kelly-weighted position sizes for each asset.

Sector ETF Hedging

Maps individual stock holdings to sector ETFs and constructs hedged portfolios to isolate stock-specific alpha from sector beta.

Qullamaggie Setup Detection

Identifies Breakout, Episodic Pivot, and Parabolic setups using pattern recognition, combined with ML scoring for position sizing.