How PatternRank builds rankings without a tip desk

Most stock ranking products ask for trust. PatternRank is built so you can inspect the process: how events are defined, how models are trained without peeking into the future, how daily shortlists are scored, and what completed outcomes looked like after the fact—including losers.

This note is a plain-language methodology brief for researchers. It is not investment advice, and it is not a claim of guaranteed edge.

Why methodology is the product

A ranked list is only useful if you can answer four questions:

1. What outcome definition was the model trained to recognize?
2. Did validation respect time order?
3. How is the live universe filtered?
4. Are completed outcomes published for the full sample, or only for winners?

If those answers are vague, you are looking at marketing. PatternRank’s public surfaces exist so you can check them yourself: the track record, sample completed rankings, and research notes.

The research loop

PatternRank’s research system runs a fixed loop. Humans design the loop. Machines execute it at scale.

market data
    → define historical events with rules
    → train models with time-aware validation
    → stress-test the process on held-out history
    → score today’s filtered universe
    → later, attach completed outcomes

There is no daily step that says “open a meeting and pick five names we like.”

1) Labels come from rules, not opinions

Historical labels are created by applying explicit event rules to price and volume history over a fixed review horizon (for the live USA shortlist: a 7-day research horizon under an ignition-style event definition).

In plain terms:

• A positive is a past setup that met the rule set for the event definition.
• Negatives are systematically sampled non-events so the model learns “ordinary days” as well as rare events. Research training commonly uses about three negatives per positive.
• Labels can be built on a liquidity-filtered universe (easy-to-borrow / more conservative names) or on a broader liquid US equity set—as two separate research universes, not as discretionary exceptions.
• Label and train windows leave a safety lag after the latest calendar day so the event horizon can complete. That is an anti-leakage choice: incomplete futures do not get treated as finished outcomes.

Analysts do not hand-select which tickers become training positives. The rules run the same way across symbols and years.

2) Training respects the calendar

Models are trained on features computed around those labeled events. Validation is walk-forward in time, not random shuffle. Folds move forward through history so the model is judged on later periods than it trained on.

A purge buffer around the horizon reduces leakage between train and validation windows. Production research maintains separate model families for the liquidity-filtered universe and the broader liquid universe, under the same event philosophy, so one set of weights is not casually mixed into the other.

Humans choose the training recipe. Humans do not sit inside the fold and promote favorite tickers.

3) Backtests stress the process

Before relying on a model family, research runs deterministic historical simulations that reuse the same scoring logic used for daily rankings. Capital is treated in a research-honest way for multi-day horizons (overlapping positions over a short window), which is a process stress test—not a live brokerage statement and not a promise of future returns.

The point of backtests here is discipline: same rules, same scoring path, reproducible seeds—not a polished highlight reel.

4) Daily rankings score the universe

On each US market day (after the close, on sessions when the market is open), the production pipeline:

1. Refreshes the instrument universe
2. Updates prices and volumes
3. Recomputes the technical inputs the models expect
4. Refreshes the liquidity / borrowability filter when configured
5. Scores both universes (liquidity-filtered and broader liquid) under the same 7-day ignition-style policy
6. Ranks names by model score and probability
7. Applies a uniform confidence floor (production currently uses a 62 minimum probability threshold)
8. Publishes the ranked research queue
9. When the 7-day horizon matures, attaches actual outcomes and updates the aggregate track record

The shortlist is a sorted model output, not a curated tip sheet. Feature contributions can be shown per name so a researcher can see what the model used—still model-derived, not a human rewrite of the rank.

“No human bias” — what we mean (and what we don’t)

It would be dishonest to claim models are free of all statistical or data bias. Humans designed:

• event definitions
• feature families
• universe filters
• confidence thresholds
• train / validation recipes

What PatternRank does not do:

• A discretionary desk choosing today’s shortlist
• Hand-editing ranks after the model runs
• Publishing only winners on the public track record
• Treating completed outcomes as optional decoration

The accurate statement is:

Human-designed research system. Machine-executed ranking.
Rules and models apply consistently. The daily list is not a human stock-picking session.

That is the property that matters for researchers who are tired of tip culture.

Two universes, same discipline

Filters are configuration. They are not “we felt like including this ticker today.”

How to audit PatternRank yourself

1. Track record — full-sample windows, including losers: patternrank.io/track-record
2. Samples — completed rankings with measured follow-through: patternrank.io/sample
3. Research notes — process write-ups: patternrank.io/research
4. Live shortlist — current rankings sit behind trial/subscription; history stays open so you can evaluate the process first

If a ranking product cannot show full-sample completed outcomes, treat the marketing copy with suspicion.

What this note is not

• A guarantee of future performance
• A claim of zero statistical bias in the abstract ML sense
• Investment advice, a recommendation, or a solicitation to buy or sell any security

PatternRank is an educational analytics platform. Rankings are for research and informational use only and are not investment advice. Past performance does not guarantee future results.