/stock-screening

name: stock-screening
description: Use when building a screen to find candidates. Covers screen design, criteria that actually discriminate, avoiding overfitting, survivorship bias, and turning a screen's output into a shortlist rather than a shopping list. Trigger on "build a stock screener", "my screen returns hundreds of stocks", "find stocks that match these criteria". Do not use for testing whether the screen's picks would have made money — see the backtesting skill.
metadata:
  category: finance
  version: 1.0.0
  tags: [screening, filters, factors, selection, bias]

Stock Screening

Purpose

Build a screen that narrows a universe to a reviewable shortlist. A screen is a filter, not a decision — its output is a list of things to look at, and treating it as a list of things to buy is the fastest way to lose money with a spreadsheet.

When to Use

• Building a systematic screen for candidates.
• A screen that returns hundreds of results, or none.
• Evaluating whether a screen's criteria actually discriminate.
• Reviewing a screen that historically performed well and now does not.

Capabilities

• Screen design: fundamental, technical, and combined criteria.
• Criteria selection and threshold setting.
• Overfitting detection.
• Survivorship and look-ahead bias avoidance.
• Ranking and shortlisting.

Inputs

• The universe, and how it is defined.
• The characteristics being sought, and why they should matter.
• Historical data, if the screen is to be validated.

Outputs

• A screen that returns a reviewable number of names — typically 10 to 40.
• Each criterion justified by a mechanism, not by backtest fit.
• A ranked shortlist for manual review.

Workflow

1. **Define the universe first** — Liquidity and market-capitalization floors. A screen that returns illiquid microcaps you cannot actually trade is returning noise. 2. **Choose criteria with a mechanism** — Each filter should have a reason it should work, stated before you test it. "It scored well in the backtest" is not a mechanism; it is a warning sign. 3. **Use few criteria** — Three to six. Each additional filter cuts the result set and increases the chance you have fitted the screen to the past rather than to a real effect. 4. **Set thresholds at round, defensible numbers** — A revenue growth threshold of 23.7% is fitted. 20% is a judgment. The former will not survive out of sample. 5. **Check the biases** — Does the historical universe include companies that no longer exist? If not, the backtest is measuring the performance of survivors, which is not a strategy you could have run. 6. **Rank and review manually** — The screen produces candidates. A human — or a careful, separate analysis — decides. The screen does not know why a company is cheap.

Best Practices

• A screen that returns 300 names has not screened. A screen that returns zero has been over-fitted. Aim for a number you will actually review.
• Every criterion should have a mechanism you can articulate before testing. Criteria discovered by searching the space of possible criteria are discovered by fitting noise, and there is a great deal of noise.
• Survivorship bias makes almost every naive backtest look good. A universe of "today's S&P 500 constituents, historically" excludes every company that failed, and that is precisely the population the screen needed to avoid.
• Look-ahead bias is subtler: using a fiscal-year figure on the day the year ended, rather than the day it was reported, gives the screen information nobody had. Use point-in-time data.
• Value screens find cheap stocks, and stocks are usually cheap for a reason. The screen finds the candidates; the work of determining whether the reason is fatal is not done by the screen.
• Do not re-optimize a screen after a period of poor performance. That fits it to the most recent regime, which is exactly the regime least likely to persist.

Examples

**A screen with few criteria, each with a mechanism:**

def momentum_quality_screen(universe: pd.DataFrame) -> pd.DataFrame:
    """Each filter has a stated mechanism. None was found by searching.

    Mechanisms:
      - Liquidity: we must be able to enter and exit without moving the price.
      - Relative strength: momentum is one of the most durable documented
        anomalies. Persistence over 6-12 months is the effect.
      - Trend structure: we do not want falling knives. Above a rising 200-day.
      - Profitability: quality screens out the low-priced momentum that is
        momentum toward zero.
      - Earnings growth: confirms the price move has a fundamental driver.
    """
    return (
        universe
        # Liquidity: tradable, not theoretical.
        .query("market_cap_usd > 2e9 and avg_dollar_volume_50d > 20e6")

        # Relative strength: top quintile over 6 months, versus the benchmark.
        .query("rs_126d > rs_126d.quantile(0.80)")

        # Trend: above a rising 200-day. No falling knives.
        .query("close > ma200 and ma200_slope_20d > 0")

        # Quality: profitable. Filters momentum-toward-zero.
        .query("roe > 0.15 and net_margin > 0.05")

        # Fundamental confirmation of the price move.
        .query("eps_growth_yoy > 0.20 and revenue_growth_yoy > 0.10")

        .sort_values("rs_126d", ascending=False)
        .head(25)                     # a reviewable shortlist, not a portfolio
    )

# Five criteria. Round thresholds. Each defensible without reference to a
# backtest. It will not be the best-performing screen in a backtest, and it is
# far more likely to work out of sample than one that is.

**The bias that makes a backtest lie:**

# WRONG: today's index constituents, tested historically. Every company that
# was delisted, acquired, or went to zero is absent. The screen appears to
# have avoided them; in reality it was never offered them.
universe = current_sp500_members()
backtest(screen, universe, start="2010-01-01")   # results are fiction

#