Quantitative Screening: A Guide to Data-Driven Investing

You open your terminal or screening platform before the market, and the same problem shows up again. Earnings headlines are noisy. Analyst notes conflict with each other. Social feeds are full of conviction with very little process. By lunch, you've seen more opinions than usable inputs.

That's where quantitative screening earns its keep.

A good screen doesn't predict the future. It reduces the search space. It takes a market that's too large to evaluate stock by stock and turns it into a smaller list that already matches your rules. That shift matters because most investors don't fail from lack of information. They fail from lack of filtering.

In practice, the edge isn't in building the most exotic model. It's in building a workflow you can run repeatedly without changing the rules every time the tape gets uncomfortable. That means combining standard financial data with signals that contribute fresh information. Insider activity is one of the few alternative data sets that can do that when handled properly.

The raw filing is rarely the edge. The edge comes from deciding which filings are routine and which ones deserve attention.

The useful version of quantitative screening sits between academic factor research and discretionary stock picking. It's systematic enough to keep you honest, but flexible enough to incorporate real-world signals like executive buying, repeated accumulation, and unusual trade size. That's the gap many investors never quite bridge. They either stay fully narrative-driven or they overbuild a model they can't explain.

Beyond the Hype Finding Your Edge with Data

Most investors have lived the same day. A stock gaps up on a headline, two analysts upgrade it, a few large accounts on social media call it early, and by the close the whole setup feels obvious. Then a week later the move fades, and you realize you never had a process. You had a stream.

Quantitative screening is the antidote to that behavior. It's a disciplined way to tell the market, “Show me only the names that fit my criteria.” Instead of reacting to whatever is loudest, you start with a thesis and let data narrow the field.

What this looks like in real life

Say you care about companies with improving operating momentum, clean enough balance sheets, and some sign that management is acting with conviction. Without a screen, you're reading filings, skimming transcripts, and chasing after names already moving. With a screen, you're reviewing a shortlist that already meets your baseline.

That changes the job.

You stop asking, “What stock should I look at today?” and start asking, “Which of these prequalified names deserves deeper work?” That's a much better question.

Where edge actually comes from

The phrase “data-driven investing” gets abused because people treat data as if volume alone creates insight. It doesn't. Useful screening comes from three things working together:

• A clear universe: You need to know which securities belong in the opportunity set and which don't.
• Rules that reflect a thesis: Every filter should exist for a reason, not because it looked good in a backtest.
• A review process after the screen: The output is a candidate list, not an automatic buy list.

Practical rule: A screen should remove work, not create a new pile of false precision.

The biggest payoff is emotional. When you already know what qualifies, it's harder to chase stories that don't fit. You can still change your view, but now you have to do it consciously.

Understanding Quantitative Screening

A workable screen starts with a practical problem. You have a large universe, limited time, and more possible signals than you can test responsibly. Quantitative screening is the rule-based process that cuts that universe down to a reviewable set of candidates using observable data.

The point is not elegance. The point is throughput with discipline.

In practice, screening works as a funnel. Early filters deal with implementation risk, such as liquidity, listing status, market cap, corporate actions, or accounting situations that can distort comparisons. Later filters focus on the traits tied to the thesis. That might mean quality, valuation, revisions, price strength, or a narrower event signal such as insider buying.

The funnel matters more than the clever factor

A lot of weak screens fail before the interesting part. They add nuanced ranking logic on top of a bad starting universe. If microcaps with sporadic volume, stale fundamentals, or one-off accounting events are mixed in with clean comparables, the output looks precise and behaves badly.

That is why the first filters often look boring. Average daily dollar volume. Minimum price. Exchange eligibility. Sector exclusions if the accounting is not comparable. I would rather spend effort getting those decisions right than adding a fifth decimal place to a composite score that sits on top of noisy inputs.

The transition from theory to implementation presents challenges. Academic research often starts with cleaner datasets than investors use in practice. A live screen has to survive filing delays, restatements, survivorship issues, and names that technically pass but are hard to own in size.

What quantitative screening actually does

Quantitative screening standardizes comparison. It applies the same rules to every name in the universe and produces a smaller set worth reviewing. That consistency matters because it keeps the process tied to the thesis instead of the news cycle.

It also has limits. A screen can flag unusual insider accumulation, accelerating margins, or improving price action. It cannot tell you whether the buyer is signaling conviction or satisfying optics, whether margins improved because of a durable change or a temporary mix shift, or whether the stock already reflects the signal.

The statistical logic is straightforward. In a normal distribution, about 68% of observations fall within one standard deviation of the mean, as described in George Mason University's overview of quantitative history and statistical reasoning. In screening, that idea helps separate routine behavior from observations that deserve attention. That is especially useful with alternative data. Insider-trading datasets contain a steady flow of SEC Form 4 activity, and a usable process ranks transactions by context and abnormality rather than treating every filing as equally informative.

Why simpler screens often hold up better

Simple does not mean naive. It means each rule earns its place.

A practical screen usually follows three layers:

• Universe rules: Remove securities that create trading, data, or comparability problems.
• Core thesis filters: Keep the variables that express the main idea clearly.
• High-information overlays: Add narrower signals, including alternative data like insider activity, where they improve selection rather than add noise.

That last layer is where many investors overreach. Insider buying is a good example. A raw filing count is weak. Context improves it. Purchase size relative to salary, cluster buying across executives, trade history, and the company's recent operating setup all matter. The edge comes from combining the event with a clean base screen, not from treating one data source as a shortcut.

If a filter cannot be explained in one sentence, it usually does not belong in the first version of the model.

Comparing Quantitative Screening Methods

A practical screen usually fails in one of two ways. It is too static, so it misses change, or it is too reactive, so it chases noise. The choice of screening method decides which mistake you are more likely to make.

Most usable workflows fall into three groups: factor screens, statistical filters, and machine-learning models. They overlap, but they do different jobs. The better question is not which one is best in theory. It is which one matches the signal you are trying to capture, the quality of your data, and the amount of validation work you can support.

Side by side trade-offs

Methodology	Core Logic	Example Criteria	Best For	Complexity
Factor screening	Select stocks with predefined characteristics	valuation, profitability, balance sheet quality, price strength	Building a stable, repeatable shortlist	Low to moderate
Statistical filtering	Flag observations that deviate from their own history or peer norms	unusual volume, abnormal price moves, outlier insider activity	Event-driven idea generation	Moderate
Machine-learning approaches	Learn patterns from many variables and their interactions	classification models using market and company features	Complex signal discovery and ranking	High

Factor screening

Factor screening is still the right starting point for many investors because it is transparent, cheap to maintain, and hard to fool yourself with. You define the traits that matter, apply them consistently, and get a list that reflects the thesis instead of your latest narrative.

That works well for stable ideas. Quality, value, capital discipline, earnings revisions, and balance-sheet strength all fit this format. If the objective is to narrow a large universe into something investable, factor screens do that cleanly.

The weakness is also obvious. A factor screen usually treats the world as slow-moving. It tells you which companies fit a profile, not whether something important changed last week. That matters if you want to blend traditional financial data with alternative signals. Insider buying, for example, rarely works well as a standalone factor. It tends to work better as an overlay on top of a sound base universe, where the event adds information instead of substituting for it.

Statistical filtering

Statistical filtering focuses on deviation. Instead of asking whether a stock looks attractive on average, it asks whether current behavior is unusual enough to justify attention.

That makes it a better fit for event-driven work. Insider activity is a good example. A single Form 4 filing often means little in isolation. The signal improves when the trade is large relative to the insider's prior behavior, appears across multiple executives, or occurs after a period of business stress when buying carries more informational weight.

In practice, this method is often the bridge between academic signal design and actual screening workflow. Academic work tends to define abnormality cleanly. Real implementation requires more judgment. You have to decide the baseline period, the peer group, the threshold for an outlier, and the conditions that should suppress false positives, such as compensation-driven trades or automatic selling plans.

A statistical filter helps rank what deserves review first. It does not remove the need for review.

Machine-learning approaches

Machine learning is useful when interactions matter more than any single rule. That is common once you combine accounting variables, market behavior, and alternative data with different update frequencies and noise profiles. A hand-built model may miss those interactions. An ML model can capture them.

The trade-off is operational, not philosophical. ML raises the standard for data cleaning, feature design, timestamp control, and validation. If those pieces are weak, the model will look smart in sample and disappoint in live use.

This is also where many screening projects become less practical than they appear on paper. A model trained on insider-related features, price action, and company fundamentals can rank opportunities well, but only if the underlying event data is standardized properly and aligned to what was knowable at the time. Filing dates, transaction codes, amended disclosures, and stale fundamentals all matter. Ignore those details and the model learns artifacts instead of behavior.

For most investors, ML is most useful after the simpler workflow already works. First build a factor screen that defines the investable universe. Then add statistical filters that identify unusual events, including contextual insider signals. Only then does it make sense to test whether a model can improve ranking inside that candidate set.

Which method fits which investor

A practical rule set is straightforward:

• Use factor screening to maintain a disciplined watchlist or portfolio universe.
• Use statistical filtering to catch fresh events, abnormal behavior, or alternative-data signals that require context.
• Use machine learning if you have enough clean history, enough features, and enough validation discipline to support it.

In many real workflows, the best answer is a hybrid. Factor screens do the first cut. Statistical filters decide what deserves immediate work. Machine learning, if used at all, ranks the survivors. That sequence usually holds up better than trying to force one method to do every job.

The Data Sources That Fuel Your Screens

A screen usually breaks at the data layer before it breaks at the model layer. The problem is rarely a lack of inputs. It is mixing sources with different update cycles, different error modes, and different uses inside the workflow.

Traditional data does the baseline work

Financial statement data sets the base rate. It tells you whether a company clears the minimum standard for business quality, balance sheet strength, and cash generation. That is the right place to answer slow-moving questions such as whether margins are holding, whether indebtedness is manageable, and whether returns on capital are improving or fading.

Market data serves a different purpose. Price, volume, volatility, and liquidity help determine whether a name is tradeable and whether the market is already reacting to new information. That matters because an idea can test well in a spreadsheet and still fail in live implementation if spreads are wide, turnover is thin, or the stock gaps on every catalyst.

Those two feeds are the foundation. They are also widely available, which means they rarely provide much edge by themselves.

Alternative data adds timing and context

The practical value of alternative data is not novelty. It is that some data sets capture behavior before it shows up cleanly in quarterly fundamentals. Insider trading data is a good example, particularly if the screen isolates discretionary open-market buying instead of treating every insider filing as informative.

That distinction matters in practice. Raw Form 4 data includes grants, option exercises, automatic sales under 10b5-1 plans, indirect ownership changes, and amendments. A naive screen will count all of it and produce a noisy watchlist. A usable screen maps transaction codes correctly, ties each filing to the actual filing date available to the market, and separates routine compensation activity from voluntary capital commitment.

What insider data can add

Insider data is useful because it records action. Executives can sound optimistic on a call for many reasons. Open-market purchases require them to commit personal capital, and that changes how the signal should be weighted inside a screening workflow.

Useful patterns often include:

• Cluster buying: Multiple insiders buying within a short window can carry more weight than a single purchase.
• Repeated accumulation: Several purchases over time often matter more than a one-off trade sized for optics.
• Behavioral change: A first discretionary buy after a long inactive period can be more informative than another routine transaction from an already active insider.

The trade-off is data cleaning cost. Alternative data can improve a screen, but only if the workflow handles filing quirks, amended disclosures, role changes, stale identifiers, and event timing correctly. That is the gap between academic intuition and a screen you can run every week without flooding yourself with false positives.

An Example Quantitative Screening Workflow

Monday morning, the screen throws up 180 names. By Tuesday, half of them are junk for your mandate, another chunk fail a basic liquidity check, and several looked interesting only because the insider data was coded too loosely. A useful workflow prevents that. It narrows the list in an order that matches how capital gets deployed in practice.

Start with a thesis and build the screen around it

Assume the mandate is conviction growth. The target is a business that still shows operating strength, where management behavior adds confirmation instead of noise.

That usually points to a simple sequence, not a sprawling factor stack. Start with investability, then business quality, then market context, then alternative data that helps rank the survivors.

A practical funnel

1. Define the universe
   Set the universe to names you can own. For a U.S. equity strategy, that often means listed common stocks only, while excluding ADRs, closed-end funds, SPAC remnants, preferreds, and other structures that create messy comparability or different reporting behavior.

2. Remove implementation problems early
   Apply liquidity and price filters before anything else. Many backtests, lacking these initial steps, often present skewed results. A signal can look strong in microcaps and still be unusable once you account for spread, position limits, and the fact that you may not want event-driven exposure in a name trading a few hundred thousand dollars a day.

3. Apply business-quality rules
   Keep this part tight. For a growth screen, I would usually want a small set of variables that capture revenue traction, margin direction, and balance-sheet resilience. The point is not to describe the whole company with ratios. The point is to remove obvious weak candidates while preserving a list that still reflects the original thesis.

4. Add market-based context
   Decide whether price action is a confirmation filter or something you want to avoid until later. Both choices can work. Requiring relative strength tends to improve trend alignment but can push you toward crowded entries. Ignoring price can surface names earlier, but it also raises the number of fundamental stories that never convert into investable setups.

5. Add the conviction layer with insider activity
   Here, theory meets workflow. Academic intuition says insider buying can contain information. In practice, the signal only helps if you standardize it well enough to run every week without drowning in false positives.

   A usable example is a cluster-buy rule set: at least 3 unique insiders, open-market purchases only using SEC Form 4 transaction code P, a rolling 15-day acute window, and a minimum individual purchase value above $25,000. If 5 or more insiders meet those conditions, the event is treated as high conviction, as described in Market Triage's description of insider trading signals.

Why this last layer earns its place

Single-insider purchases are easy to overread. A small buy can be symbolic, role-related, or too small to matter economically. A properly filtered cluster is harder to dismiss because it asks for multiple people to commit personal capital within a tight period and under comparable transaction rules.

That does not make insider buying a standalone strategy.

It makes it a ranking tool. After the liquidity, quality, and market filters have done their job, insider activity can help decide which names deserve research time first. That is the practical bridge between the academic idea and a screen an analyst can maintain.

What happens after the screen

The output should be short enough to review one by one. If the list is still too long, the workflow is not selective enough.

The next step is manual review:

• Check filing context: Confirm the trades are discretionary open-market purchases and not artifacts of amendments, family entities, or unusual ownership structures.
• Review the event calendar: Earnings dates, guidance changes, financings, and M&A rumors can all change how the signal should be read.
• Test the business case: Make sure the operating picture, valuation, and insider behavior point in the same direction.

That final review is where a screen becomes investable. The model narrows the field. Judgment decides whether the name belongs in the portfolio or only in the watchlist.

How to Backtest and Validate Your Screens

A screen ranks 20 names beautifully in a spreadsheet. Six months later, the live basket underperforms because half the signals relied on filing timestamps, revised fundamentals, or liquidity assumptions you could not have traded in real time. That gap between research performance and executable performance is where most screening ideas break.

Backtesting has one job. Recreate the decision process as it would have run on each date. For a practical screen, that means point-in-time data, realistic rebalancing rules, and a delay structure that respects when filings, estimates, and alternative data signals became available. Insider trading data is a good example. The event may be economically meaningful, but the test still fails if the model assumes immediate awareness when the filing would only have been visible later.

What good validation looks like

Start with the full workflow, not just the factor formula. Define the universe at each point in time. Apply the same liquidity rules you would use in live trading. Rank securities with the information available on that date. Then simulate holding periods, turnover, and transaction frictions at a level that matches the strategy you could realistically run.

That sounds basic, but it changes results fast.

A useful validation process also checks whether the screen behaves consistently across different environments. A signal that only works in one liquidity regime or one style cycle may still be interesting, but it should be treated as regime-dependent, not as a general edge. The same applies to alternative data. Insider cluster buying can add information, yet its value usually depends on context such as company size, event timing, and baseline volatility.

Metrics that matter for signal quality

Total return is only the surface layer. A workable screen also needs to answer narrower questions. Does the top decile outperform the middle of the rank order, or are results driven by a few outliers? Does the hit rate hold up after costs? Does turnover stay within the limits of the mandate?

For classification-style signals, especially event-driven ones, ranking metrics matter because the ultimate decision is often about prioritization. AUC-ROC can help assess whether the signal separates stronger candidates from weaker ones across thresholds. Precision often matters more in day-to-day use. In an insider screen, low precision means analysts spend time on names that looked interesting in the model but had little informational value in practice.

A noisy screen wastes two scarce resources. Research time and risk budget.

How to avoid fooling yourself

A few habits reduce self-deception without making the process academic for its own sake:

• Use a clean out-of-sample test: Keep a period untouched while you set the rules. If the idea only works in the sample that shaped it, the edge is probably overstated.
• Build in implementation frictions: Include realistic assumptions for slippage, spreads, and position limits, especially in smaller names where alternative data signals often look strongest.
• Check signal decay: Measure how quickly the edge fades after the screen date. That determines whether the strategy fits a weekly review, a monthly rebalance, or a shorter event-driven process.
• Review the losers by hand: Failed selections often reveal timestamp errors, hidden exposures, or a context filter that should have been there from the start.
• Run a forward period: Paper trading or a small live sleeve can expose issues that never show up in historical tests, particularly around data latency and execution quality.

I also want the candidate list to remain explainable. If the backtest improves every time another exception rule is added, the model is learning the sample instead of the market. A screen should narrow the field with repeatable logic, then leave room for judgment. That is the practical bridge between research and implementation, especially when you mix standard market data with alternative inputs such as insider activity.

Avoiding Pitfalls and Implementing Your Strategy

Most broken screening systems don't fail because the idea was irrational. They fail because the user asks the screen to do too much. A screen is a filtering tool. It's not a substitute for judgment, context, or trade construction.

The mistakes that show up repeatedly

• Over-optimization: The rules become so specific that they describe one historical sample instead of a repeatable process.
• Data snooping: You test so many combinations that something good appears by accident.
• Context blindness: You treat every passing name as equivalent even when the surrounding facts are clearly different.

The habits that make screening useful

A practical implementation style usually looks like this:

• Keep the first version simple: If you can't explain the logic in a few sentences, simplify it.
• Treat outputs as candidates: Run deeper research before committing capital.
• Review and rebalance on a schedule: Don't let random market noise dictate when you revisit the process.

The strongest quantitative screening workflows are rarely the flashiest ones. They're the ones an investor can run next month, next quarter, and next year without changing the rules because of one bad headline or one missed trade.

The screen should make you more disciplined, not more reactive.

That's its primary value. Quantitative screening gives you a repeatable way to narrow the market, focus your attention, and integrate alternative data without drowning in it. Used properly, it doesn't replace discretionary thinking. It upgrades it.

---

If insider activity is part of your process, Altymo is worth a look. It turns raw SEC Form 4 filings into usable buy and sell alerts by scanning 5,000+ filings per day and surfacing the transactions most likely to matter, including cluster buying, unusually large trades, first-time buying after inactivity, and CEO or CFO open-market purchases. For investors building a quantitative screening workflow, that can be a practical way to add an alternative data layer without manually parsing the entire filing stream.