How to compare two screen runs
Comparing two screen runs tells you what changed between two snapshots of the same thesis. Quantery sorts the diff into Added, Dropped, and Still passing, then shows score and rank movement for the names present in both runs. That turns a fresh survivor list into a set of questions you can investigate.
Run the same thesis again after new filings or market prices arrive, and a company can cross a gate, lose a point, or move within the ranking. The comparison shows the movement. It can't tell you why by itself. Keep the thesis version and run settings fixed, then inspect the values retained with each run. That narrows the question without pretending the diff preserved a historical filing trail.
When should you compare two runs?
Use a run comparison when your question is about change over time. A current scan answers, "Which companies pass these rules now?" A comparison asks, "Which part of that answer changed since the earlier run?"
A useful cadence depends on the thesis. A price-sensitive thesis can move whenever prices move. A thesis built mostly from quarterly statement values may stay stable between filings, then change when a company files a new report. There's no virtue in rerunning on a calendar just to produce activity. Pick an interval that matches the inputs you're studying.
The cleanest comparison uses the same saved thesis version and the same recorded run settings on both sides. Check the symbol or full-universe scope, any gate overrides, and the qualitative-scoring settings, including whether scoring ran and which model and cap it used. A matching version doesn't prove those settings match.
Quantery warns when thesis versions differ or when either version is unknown. Treat that warning as a stop sign for causal claims, but don't treat its absence as proof of a controlled comparison. The diff remains useful when settings differ. It just can't isolate new data as the cause.
This is the same discipline used in a good robustness test. Change one dial and rerun when you mean to test the dial. Keep every dial fixed when you mean to observe new data. Mixing the two leaves you with movement and no clean explanation.
How do you open the run comparison?
Open a thesis with two distinct recorded run IDs on different UTC dates. Its Run history lists the stored scans. Running the thesis again on the same UTC date replaces that date's earlier stored results, so two same-day scans don't produce two rows to compare. Select an older row if you want to inspect that run on its own; Back to latest returns you to the newest results.
Choose Compare, then select the two runs you want to put side by side. The Comparing runs panel replaces the ordinary survivor view with the diff. Close returns to the normal view. The Quantery documentation for scanning and run history covers the same controls.
Don't begin with runs months apart unless that span is the question. Start with adjacent runs. Fewer intervening filings and price moves make each difference easier to explain. Once you've learned what normally changes from one run to the next, widen the interval to study turnover over a longer stretch.
The panel orders the pair from older to newer. Check that direction before interpreting rank movement. A lower rank number means the company climbed within the later survivor list. It doesn't mean the underlying business improved, and it certainly doesn't establish that the market mispriced it.
What do Added, Dropped, and Still passing mean?
Added means a company passes the quantitative gate in the newer run but didn't pass it in the older one. It may have been a near-miss before. It may also have lacked a required input, fallen outside the thesis universe, or not appeared in the prior data snapshot. Added means newly present among the gate survivors. It doesn't mean newly attractive in any broader sense.
Dropped is the reverse. The company passed in the older run and doesn't pass in the newer one. That can happen because a feature deteriorated, a price-dependent ratio crossed a cutoff, a filing filled a previously missing value, or the company ceased to be eligible for the declared universe. A drop isn't a verdict on the business. It's a statement about one ruleset evaluated against a later snapshot.
Still passing contains companies that pass in both runs. For each one, Quantery shows the earlier and later composite scores, the score change, both ranks, and the rank movement. These rows often do more work than Added or Dropped. A stable membership list can hide a changing order, and the order tells you which names moved relative to the rest of the eligible group.
Remember that rank is relative. A company's score can stay fixed while its rank falls because other companies improved or entered ahead of it. Its score can decline while its rank rises if the rest of the survivor list declined further. Percentile ranks and deciles explains the same dependency: every relative measure inherits the universe around it.
Read score and rank together. Score movement points toward the company's own evaluated features. Rank movement tells you how that change sat among the other survivors. Neither line is a reason to act. Each is a clue about where to inspect next.
How do you explain why a company moved?
Start with the thesis version and the recorded run settings. If either differs, identify the changed parameter, feature, gate, universe scope, override, or scoring configuration before blaming new company data. Only matching versions and matching settings make new inputs a plausible explanation for the movement.
Next, compare what the historical runs actually retain: membership, composite score, and rank, plus the recorded feature values available in each saved result. Those values can show that a stored feature changed between runs. They don't preserve the old calculation's expanded inputs or filing provenance, so they can't establish which filing line caused the change.
The explanation drawer needs careful reading here. It can show the value recorded with the selected run beside a live value computed from today's lake. Its matched rule, expanded inputs, fiscal periods, filing dates, and source links belong to that live re-evaluation. Every number can show its work covers that current provenance layer. Don't attach it retroactively to the historical value.
Use the retained values to form the next question. A changed valuation feature may send you to the price history. A changed statement feature may send you to the filings that became available between the run dates. A membership change may send you to universe eligibility or data coverage. Read those records independently, and leave the cause unresolved when the historical evidence isn't retained. The diff finds the row; it doesn't preserve the old arithmetic.
What can a run comparison fail to show?
The comparison covers gate survivors. It isn't a full audit of every eligible company and every near-miss. A company absent from one side may have failed the gate, fallen outside the universe, or lacked data needed for evaluation. Open the run and its coverage details before assigning a single cause.
Very large results carry another boundary. The app fetches and compares the first 1,000 rows from each run, and the panel says when that limit was reached. That figure comes from the product documentation. In a truncated comparison, a name missing from the displayed diff may sit below the fetched portion. Don't turn absence from a partial view into a claim about absence from the complete run.
A pair of runs also doesn't measure portfolio turnover. Added and Dropped describe membership among thesis survivors. A portfolio test may apply a holding count, rebalance schedule, or weighting rule that selects a smaller set and trades on different dates. If the research question concerns performance or trading, use a backtest. Every backtest result is hypothetical and excludes costs, and a good benchmark must match the thesis's opportunity set.
Nor does repeated membership validate a thesis. A bad rule can produce a stable list. Stability tells you that the result isn't twitching under the observed data changes. It says nothing about whether the selected feature predicts the outcome you care about.
How do you turn the diff into better research?
Keep a small change log beside the thesis. For each comparison, record the two run dates, thesis version, run settings, and the movements worth investigating. Write a cause only when independent historical records support it. "The recorded feature fell below the gate between these runs" stays within the diff. "The latest filing caused the drop" needs the intervening filing and calculation to prove it.
After several comparisons, look for patterns in the rules rather than stories about individual companies. If one criterion drives most of the turnover, test whether it needs a longer window or a scoring band instead of a hard gate. If ranks jump while scores barely move, the universe or cross-sectional distribution may be doing the work. If missing values create repeated entries and exits, fix the null policy before judging the signal.
That's where adaptability matters. The comparison isn't asking you to trust a changing list. It shows how your own equations respond as their inputs change. You can inspect the rule, soften a brittle cutoff, separate a price signal from a filing signal, and rerun the same test.
Use the diff as a queue for investigation. Added tells you what crossed in. Dropped tells you what crossed out. Still passing shows whether the center held or rearranged itself. Then open the numbers and do the part no label can do for you: decide whether the movement belongs to the business, the market, the data, or the rules you wrote.
Research tooling, not investment advice. Nothing here is a recommendation to buy, sell, or hold any security. Screens, scores, and backtests are informational only; backtested results are hypothetical, exclude costs such as commissions and slippage, and do not guarantee future results. Verify against primary filings and make your own decisions.
Want to try this on your own rules? Quantery is free for 14 days: the full app, no card required.
← All articles