📈 How It Works

Every trading day, our pipeline screens ~5,300 US stocks through a rigorous 5-stage process, filtering, scoring, and ranking to surface a handful of sustainable investment opportunities.

The 5-Stage Pipeline

The Scoring System

Every stock receives a combined score out of 100 from four components:

Combined = Q + E + S + B (max 100). Sub-scores visible in every report: e.g. 82/100 (Q:48 E:14 S:8 B:10)

Two Buckets

Signals are split into two buckets to match different investment profiles:

• 🚀 Growth ($5–$20): smaller-cap stocks with higher return potential. Our backtest data shows the lower-priced ranges have historically delivered the strongest returns over long holding periods.
• ⚖️ Momentum ($20–$100): established companies with confirmed uptrends. More recognisable names with strong institutional backing.

See full backtest results →

ESG/SRI Filtering

We use SEC EDGAR SIC codes to systematically exclude companies in industries that conflict with sustainable investing principles. Our exclusion list covers 11 categories (fossil fuels, tobacco, alcohol, gambling, weapons and defense, mining, banking and finance, pharmaceuticals, predatory lending, animal welfare, and private prisons), supplemented by keyword matching and a manual review list. Companies misclassified by SIC codes (e.g., tech firms under financial codes) are whitelisted after verification.

Signal Selection & Scoring Framework

Signal weights aren't guesswork. Our methodology separates two distinct problems: which signals belong in the model, and how much weight each deserves. Each active signal has a full research page showing its backtest data.

Phase 1: Signal Selection

Before a signal earns any scoring weight, it must demonstrate genuine predictive value across multiple independent tests. We use three lenses to decide whether a signal belongs in the model at all:

• Walk-forward consistency: each signal is tested independently across multiple non-overlapping out-of-sample periods spanning our 14-year backtest window. Signals that only worked in one market regime, or that fall below a minimum consistency floor, are eliminated regardless of their in-sample performance.
• Mutual Information (non-linear relevance): measures how much each signal independently predicts future returns without assuming a linear relationship. Captures patterns that correlation-based methods miss entirely. Signals with near-zero relevance are removed at this stage.
• Redundancy analysis (mRMR + Elastic Net regularisation): ensures each selected signal adds information beyond what others already capture. A signal that is highly correlated with a stronger signal is eliminated even if it looks good in isolation. Signals that collapse to zero under joint regression are removed.

Signals failing more than one of these tests are removed. A compelling walk-forward result doesn't earn inclusion if the signal proves informationally redundant.

Phase 2: Weight Optimisation

For signals that pass selection, we determine how much scoring weight each deserves:

• Regression coefficients: joint regression across all surviving signals provides initial weight direction and relative magnitude.
• Walk-forward return deltas: out-of-sample return improvement confirms the weight direction holds outside the training window.
• Information Coefficient / IC-IR: measures the correlation between each signal and actual forward returns, computed independently across each test window. The ratio of mean IC to its variability (IC-IR) rewards signals that are both predictive and consistent across different market conditions.

This two-phase approach prevents the failure mode of promoting signals that look strong in isolation but are measuring noise or information already captured by a better signal.

Combinatorial Signal Analysis

Beyond individual signals, we systematically test all two-way and three-way signal combinations (2,293 in total) to discover which pairs and trios of signals compound returns when they fire together. This uncovered our strongest confluence: stocks near their 52-week high with strong return on equity consistently outperform across all time horizons tested.

Browse the Signal Research Library →

Phase 3: Genetic Algorithm Weight Optimisation

After signal selection and individual weight derivation, we run a Genetic Algorithm (GA) optimizer that searches across the entire combinatorial space of valid weight configurations simultaneously, something grid search or manual tuning cannot do at scale.

The GA evolves a population of candidate weight sets over hundreds of generations, scoring each one against the same metric we care about: cumulative portfolio alpha vs the S&P 500 across the full 14-year backtest (not a proxy metric). Hard constraints ensure every candidate must fire across multiple non-overlapping market regimes (covering the GFC recovery, multiple bull runs, the COVID crash, the 2022 bear market, and the subsequent recovery) with a minimum frequency floor to prevent cherry-picking a small number of historically perfect stocks.

The current signal weights (R30) were GA-optimized. Key findings from that run:

• Valuation earned the most weight: the GA gave a low price/earnings-to-growth (PEG) signal, in the spirit of Peter Lynch, the single largest share of the quality score, rewarding companies whose growth is not yet reflected in their share price
• Entry timing favoured diversity: it spread the timing budget across more independent signals, adding a record-volume accumulation day and an overbought-momentum reversal, rather than leaning on any single pattern
• Quality concentrated on the strongest cross-regime trends: persistent trend signals gained weight while marginal ones were retired, keeping the quality score on measures that hold up across market cycles
• Fewer, higher-conviction picks: the optimized weights identified 56 high-conviction stocks across 14 years, tighter selectivity at higher alpha

The GA never modifies production settings directly. It outputs a human-readable suggested diff for review before any change is applied.

What We've Tried (And Removed)

Our approach is ruthlessly data-driven. Many popular technical signals were tested and removed because the data showed they didn't help, or actively hurt, long-term returns:

• Confluence signals tested: combinations of individual signals where most showed neutral or negative return deltas when paired; only the strongest confluences survived validation
• Non-differentiating signals removed: signals that fire for the vast majority of stocks provide no useful differentiation and were eliminated
• Harmful signals identified: several widely-used technical signals consistently showed negative return deltas across multiple time horizons after controlling for redundancy with stronger signals
• Failed walk-forward validation: signals that appeared promising in-sample failed to show consistent predictive power across independent out-of-sample test periods and were removed
• Signals re-evaluated across rounds: as our validation framework improves, every signal gets a fair re-evaluation. Some signals initially removed have been restored after more rigorous testing confirmed their effectiveness

Every signal earns its place through empirical performance, not tradition or popularity.

What the Validation Methods Are Optimising For

The methods work in two phases with different but complementary objectives. Walk-Forward (WF), Mutual Information (MI), Elastic Net regularisation (EN), and the Information Coefficient / Information Ratio (IC-IR) all evaluate signals against the same target: raw portfolio returns over a 7-year holding period, independent of index performance. The Genetic Algorithm (GA) takes a different objective in the final step: it directly optimises the combined weight configuration against actual portfolio alpha vs the S&P 500, across five independent temporal folds, to prevent overfitting to any single market period. Each method approaches this from a different angle:

Method	Full Name	What it optimises for
WF	Walk-Forward	Does this signal fire on stocks that returned more over the next 7 years, consistently across independent out-of-sample test windows?
MI	Mutual Information	How much does this signal independently predict 7-year returns, without assuming a linear relationship? Near-zero relevance means removal.
EN	Elastic Net regularisation	Given all signals together, what weight minimises 7-year return prediction error? Signals whose coefficient collapses to zero under joint regression are removed as redundant.
IC-IR	Information Coefficient / Information Ratio	How consistently does the signal correlate with forward returns across different test windows? High IC-IR distinguishes genuinely predictive signals from those that worked in just one market regime.
GA	Genetic Algorithm	Once signals are selected and individually weighted, a GA searches across the full combinatorial space of valid weight configurations simultaneously, scoring each against actual portfolio alpha vs the S&P 500. Hard cross-regime constraints (must fire across multiple independent time windows) prevent the GA finding configurations that only worked in one historical period.

Beyond per-signal validation, we test the whole configuration out of sample. We do not just optimise on history; we check that the optimisation itself holds up on data it never saw. The complete weight configuration is rebuilt using only the earlier years, then tested on the later years it was blind to, repeated across several split points that span market crashes, bull runs, bear markets, and recoveries. The bar it has to clear in those unseen windows: the blind-built configuration still beats the S&P 500, and still out-picks simply owning the entire screened list, in most of them. We also compare the top candidate configurations for that kind of steadiness across regimes, not just headline return.

Backtest Limitations

The 14-year backtest uses a data source that includes delisted and acquired companies, so the screened universe is not restricted to stocks still trading today. When a flagged stock stops trading before a hold period ends (acquisition, delisting, or bankruptcy), the return is recorded to the last available traded price. For acquisitions this captures the deal premium accurately. For bankruptcies, the final pre-delisting close is used, which may slightly understate the total loss compared to an eventual $0 outcome.

The practical impact is reduced by the quality scoring: the long-term debt/equity and earnings-based signals select against the overleveraged or unprofitable companies most likely to delist involuntarily. A company with negative or shrinking earnings cannot earn the valuation signal that now leads the quality model, and the debt/equity signal screens against excessive leverage. Case studies including losses are published alongside wins. Results should be treated as backtested estimates, not guarantees of future performance.

The reported alpha figures use a daily-cadence backtest, consistent with how the scanner operates in practice. Some entry signals show higher alpha at a weekly screening frequency, but their characteristics change when evaluated daily; using daily cadence avoids reporting performance that would only materialise if stocks were re-screened once per week rather than every trading day.

Evolving Methodology

We continuously rebalance and validate our scoring system. The quality score is fundamentals-first: valuation (PEG), profitability, balance sheet strength, earnings growth, and revenue momentum account for the majority of points.

The methodology keeps evolving as we gather more data, test new signals, and refine our understanding of what drives sustainable long-term returns. Walk-forward analysis across 14 years confirms that fundamental signals consistently predict long-term returns better than technical pattern signals alone. Signals are promoted or removed based strictly on empirical evidence across multiple validation methods. See the full revision history.