CompanyGraph
How predefined filters make an investment hypothesis testable without pretending that the result is a conclusion.
A screen begins with a question
“Find good stocks” is not a screen. A useful screen begins with a question such as: which companies have improving cash conversion and manageable leverage, or which businesses are growing while retaining pricing power? The criteria, thresholds, comparison universe, date, and missing-data rules then make that question repeatable.
Adding criteria narrows the set but does not automatically improve the answer. A screen for high return on equity can include heavily leveraged companies. A low price-to-book screen can return businesses whose assets no longer earn their historical returns. A growth screen can find companies that buy revenue with cash and margin. The combination is a hypothesis about what conditions should coexist, not a proof that they do.
What the data actually observes
Financial screens use reported revenue, earnings, assets, liabilities, cash flow, prices, and shares. The SEC explains that a Form 10-K contains audited statements and management discussion, while also warning readers to consider risk disclosures and non-GAAP measures. A screen can process those fields consistently, but it cannot see a customer conversation, a product defect, or a covenant negotiation that has not reached the filing. The SEC's financial-statement guide describes the source boundary.
Align dates. Annual revenue can be paired with a period-end balance sheet, while the market price may be from a different day. A trailing metric can include a business before an acquisition and after it. A screen should record the point-in-time convention rather than imply that every input describes the same condition.
Combinations can reveal a profile
Quality screens might combine profitability, cash conversion, leverage, and stability. Value screens might combine a low multiple with liquidity and improving operations. Growth screens might require revenue expansion with stable margins and cash generation. Risk screens can invert those conditions to identify fragile balance sheets.
The categories are not universal. A bank, software company, manufacturer, and early-stage biotech need different measures. An empty result may indicate that the criteria are contradictory, that the universe is small, or that the requested profile is absent at that date. A large result may mean the filters do not distinguish much.
Thresholds create hidden decisions
“ROE above 15%” treats 15.1% and 14.9% as different while treating 15.1% and 35% as equivalent. Percentile ranks avoid some cliffs but depend on the comparison set. Equal-weight composites make a mathematical claim about the relative importance of their components. Each choice should be disclosed and stress-tested.
Run sensitivity checks: vary thresholds, exclude one metric, use a different industry classification, and test the result before and after major accounting or market events. If one small choice changes the entire list, the apparent precision is fragile.
Why a passing screen is not a recommendation
A screen says that the selected observations meet the rules. It does not say that the stock is cheap, that the condition will persist, or that management can defend it. Price, competitive position, product concentration, regulation, capital needs, and customer behavior remain outside many screens.
Survivorship bias can also make a historical screen look better than it was. A current universe omits companies that failed, merged, or delisted. Backtests must define the historical universe and include delisted securities where possible. Otherwise, the method is partly evaluating the survivors it already knows.
Use exclusions as research questions
When a company fails, preserve the reason. A low cash-conversion score may reflect working-capital investment, or it may signal weak collection. High leverage may finance a durable asset base, or it may leave the company exposed to refinancing. A screen cannot decide between those explanations, but it can make them visible.
When a favored company passes, ask what the screen cannot see: customer concentration, patent expiry, product quality, employee capability, environmental liability, or a change since the last filing. A passing result should trigger primary research rather than replace it.
How to build a durable screen
Write the hypothesis and intended holding period first. Define the universe, accounting basis, dates, currency, sector treatment, and handling of missing data. Use a small number of criteria that each have a reason to be there. Keep the components visible and record changes to the method.
Validate the method on periods and companies not used to design it. Compare results with a simple baseline and test turnover, transaction costs, liquidity, and concentration. The screen is successful when it improves the quality of questions and decisions, not when it produces a reassuringly short list.