Interactive Market Data Research
Essays on signal calibration, backtesting methodology, and the data infrastructure behind Interactive Market Data's SEC analytics. Honesty-as-marketing — the signals that don't work get the same airtime as the ones that do.
-
How retail-calendar fiscal years break every off-the-shelf data pipeline
Costco, Walmart, Home Depot, Target, Best Buy, Lowe's — most American retailers use a 52- or 53-week fiscal calendar that ends in the last week of January. SEC's API and most commercial vendors return the period correctly but their `fy` field is wrong by one for the rolling-year comparatives. We caught it during our verification pass; fixing it dropped 50+ false positives. Here's the bug and a five-line helper that solves it.
-
Why our calibration framework dropped quadruple-signal stacking entirely
We backtested 4-signal stacks across our 12-signal library. Several looked like elite r/σ. All of them had n in the 10-30 range, failed out-of-sample validation, and produced cohorts where 80%+ of events came from a single sector or a single ticker. We stopped publishing quadruples. Here's what the data showed and why we don't believe it even when it looks good.
-
How we got our SEC replica to >99.99% alignment — four stacked fixes
We rebuilt our SEC replica against parquet (no ClickHouse) and the first pass showed 4,042 cells "wrong" out of 71,198 sampled. After fixing four orthogonal things — the verify tool, an XBRL display flag, retail-calendar fiscal year detection, and BS pivot tie-breaks — the residual was 4 noise rows. 99.994% SEC alignment. Here's the sequence and why each fix mattered.
-
Real-dollar-flow as the leading indicator: a new lens on equity capital
Apple at a $4T market cap doesn't mean $4T moved today; it means the last marginal share traded at that price. The capital that actually reallocated is |price-change × volume|, summed daily. SLV's signed real-dollar-flow has been -$10.6B YTD even as silver rallied — informed money has been distributing throughout. Here's how RDF works, what it shows that market cap can't, and how to integrate it into a fundamental research workflow.
-
Macro conditioning doubles your alpha (or wipes it out)
capex_spike short returns -4.21% over 20 days during a flat yield curve — and roughly zero during a steep curve. zombie_alert is a short during flat curves and a long during steep ones. We backed this out across 30K+ events. Most signals don't work in all regimes; the ones that do are the ones publishing single-number calibrations are quietly hiding their mediocre regime.
-
Stacking is real, but only honestly: pair vs. triple vs. single signal calibration
Single-signal calibration shows capex_spike short returns -1.91% over 20 days. Pair it with fcf_turn_negative and you get -2.65%. Add fcf_ni_divergence and the triple shows -3.78% with r/σ of -0.39. But the academic literature on signal stacking is mostly snake oil — what's actually robust, what's overfit, and how to tell the difference.
-
Why your favorite fund's data has the same negating-flag bug we found
When we ran our SEC replica against ground truth, 653 cells looked wrong. Investigation: the bug was in our display layer applying XBRL's `negating` attribute to stored numerical values. SEC's own companyfacts API returns raw values — `negating` is purely a render-time layout hint. The same bug is present in every fundamental data vendor we sampled. Here's how to find it in your own pipeline.
-
How a phantom signal almost shipped — and how the validation killed it
We discovered a triple-signal stack that looked like institutional-grade short alpha (n=72, mean -5.89% over 20 days). Validation revealed the entire signal was a 64-fold duplication of a single ticker. The deeper finding after the fix: at 252-day horizons, multi-stress signal stacks calibrate as deep-value LONGS, not shorts. Survivor bias rules the long horizons.