Topic
Quantitative research
Point-in-time data, back-tests and systematic signals.
- Testing whether a job postings indicator adds to official statisticsTwo models forecast one official figure: its own past, then the same model with postings added. The design, the vintage tables it needs, and a worked example in which a lower error still fails its interval.
- How to seasonally adjust a job postings time seriesJob postings follow the calendar as well as demand. What the official seasonal adjustment methods ask of a series, what they cannot remove, and two comparisons that need no model, worked through on an illustrative weekly series.
- How to build a sector hiring index from company job postingsOne large employer can decide a sector index built by summing its postings. Here is how to weight, cap and bound company-level hiring, with SQL on daily data and an illustrative example.
- How company events are dated: occurrence date or publication timeAn event has the day it happened, the time the company published it, the time a filing was accepted and the time a feed first saw it. Which one to use for a test or an alert, and how to check each at its source.
- Measuring signal decay in an alternative dataset as funds crowd inSplit the history into periods, measure the rank correlation in each, fit the fall, then test for a break before calling it crowding. The statistics come from NIST publications, with an illustrative example worked through.
- Do job postings predict company revenue growth? What official studies showA 2013 Reserve Bank of Australia study ties job advertisements to employment a quarter or two ahead, for a whole economy. For one company's revenue the answer is a test you run yourself, and this post sets it out.
- How to link private companies and brands to a listed parent companyWho owns whom is stated in filed records, and each defines a parent differently. This guide compares them, builds a dated bridge and rolls brand observations up to the parent's security without counting anything twice.
- Same-store cohorts: how the weekly market series stay comparableCoverage of a company index expands continuously, and a raw count would carry that expansion into every trend. Here is the cohort rule that keeps it out, and the arithmetic of every measure it feeds.
- Why we write each table once: point-in-time data by designA table corrected in place shows what is known today, not what was known then. The five rules that keep Fokals tables as first written, what they guarantee and how to check them.
- Building a point-in-time company dataset for back-testsA back-test is only as honest as the dates in its data. This guide shows how a dated, write-once record lets you key company signals by when they were known and query as of a date.
- Building thematic baskets from technology adoption dataA basket of adopters is only as good as its membership rule. Here is how to build one from website technology data, rebalance on dated events and test it.
- Mapping alternative data to a security masterJoining company data to securities is a many-to-many problem. This guide covers the identifiers every listed company carries, the join order, parents and brands, delistings and check digits.
- Training and evaluating models on dated company signalsA model on company signals is only as honest as its feature table. This guide builds one as of each date, splits by time and names the properties of the record that keep it safe.
- Look-ahead biasA back-test that uses information from after the date it simulates looks better than any real decision could have been. How it enters a test and how to find it.
- Point-in-time dataData stored as it stood on each past date, with no later corrections. How the two dates of a record work, an as-of query and the mistakes that break a back-test.
- Same-store cohortCompare only the companies present for the whole window, so wider coverage does not look like market growth. The method, a worked example and where it is used.
- Survivorship biasStudying only the companies that lasted makes past results look better than they were. Where the error appears, how to test for it and how to keep delisted names.