Field notes on alternative data.
What the filings actually say, how the pipelines are built, and which signals survive contact with a backtest.
Do WARN notices really give 60 days' warning?
On six states whose fields we verified, the median WARN notice is filed 60 days ahead and 63.0% since 2021 give 60 days or more. Against 80 notices that overlap an 8-K Item 2.05, the notice came first only 32 times.
Read postFails-to-deliver, the threshold list, and the short-squeeze folklore
After a symbol joins the Reg SHO threshold list its short interest falls a median 12.6 points more than matched controls, but nearly all of that gap comes from listings that follow a doubling of short interest.
Request logging is lossy on purpose
Usage rows go through a bounded in-memory queue and can be dropped; quotas run on Redis counters and cannot. Why the two are separate, and what it means for the numbers you see.
Congress trades. We time-stamped the disclosures.
Median disclosure lag for House stock trades is 15 days, the 90th percentile is 35, and 7.9% arrive after the 45-day deadline. Key your backtest on the disclosure date.
Indexing is cheap, fetching is expensive
Finding out a filing exists scales with filings; reading it scales with symbols and bytes. That asymmetry decided which SEC feeds are broad and which are gated.
Form 144 says they're going to sell. Form 4 says whether they did.
Across 96,851 proposed insider sales since 2024, 41.3% are followed by a Form 4 sale in the next 30 days, and nearly all of those on the same day.
CUSIP, CIK, ticker, LEI: four identifiers, none of them primary keys
Every financial dataset is keyed on something that changes. Here is what each identifier actually names, how each one breaks, and what we key on instead.
Nobody files an NT 10-K for a good reason
A Form 12b-25 is an issuer admitting the numbers are not ready. Across 30,754 notices since 2015, here is what actually happens next.
Wikipedia pageviews are a retail attention index
A free daily count of who is looking a company up. It forecasts volatility, not direction — and the baseline is the entire signal.
We deleted most of our MCP tools and the agent got better
One tool per endpoint does not scale past about a dozen. Here is what we replaced it with, and what each decision cost.
Why we built HedgeFriend
Public-domain data is free, but it is not cheap. Here is the friction we set out to delete.
Reading Form 4: what insider filings do and don't tell you
Insider buys are a noisy signal until you strip out the scheduled sales. A short field guide.
Your macro backtest is probably cheating
Most macro series are revised after publication. If your history stores only the latest value, you are trading on numbers nobody had.