How the numbers are made

Every result on Underfolio is counted from market history. This page says what that history is, what was adjusted, what was assumed, and where the edges are. Each result page links to the part that applies to it.

Coverage

• Stocks: every US-listed common stock in CRSP from January 1968, plus supplemental daily histories back to January 1962 for a small number of names, plus newer listings from other sources. About 45,500 series, roughly 28,000 of which have stopped trading. The Graveyard is the 5,000 biggest of those by peak market value.
• Dead companies count. A stock that ended is in every scoreboard and rule test; nothing is dropped because it failed later (no survivorship bias). A stock that stopped trading is not necessarily a failed company: most were bought, merged or taken private, and each obituary says which where our delisting record knows.
• Market data: the S&P 500 (the Dow before it existed), gold, oil, the dollar, Treasury yields and other series, some back to the 1880s. Each Lab answer says the first year it could see for the situation asked.
• Economic data: US series (CPI, unemployment, payrolls, GDP, Fed funds and the rest) are used as first published, on the day they came out, so a situation is judged with the numbers people had that day. Euro area, UK and Japan series are revised figures with an estimated release day; answers that use them say so.

Prices, adjustments and returns

Prices through 2026-10-01 in the stock collection. Individual stocks, indices and economic series can end earlier; each chart shows its own last date.

• Charts and rule tests use prices adjusted for splits and dividends (total-return adjusted), so a move across a split is a real move.
• Obituaries compare a split-adjusted high with the last trade; the peak market value is as it was on the day. "Return over the days in our data" compounds daily total returns between the first and last day we have; it does not include what holders received after the last trade (a merger payout, or nothing in a bankruptcy) and does not reach back before our first day.
• Before 2014 most dead companies have only a daily close, not an open, high and low. Those days draw as a line and are scored apart: the pattern scoreboard's headline reads real candles only, and a custom rule that reads the range (high, low, ATR and friends) skips close-only days rather than treating them as zero-range bars.
• The last numbers a company filed (on obituaries from 2009 on) come from the SEC's company facts: its last annual or quarterly report filed on or before its last trade, in US dollars as reported. Nothing filed later is used. Source: SEC EDGAR.
• Delisting returns are not in our copy of CRSP. A stand-in is used where one is needed and the page says so.
• Market value and shares after 2024 are estimated for some names.
• "What if I had invested" buys at the close of a stock's first trading day of the year and puts dividends back in, with no costs and no tax. Dividends are counted to the end of 2024: from 2025 on our prices are adjusted for splits only, so every total return that runs past 2024 leaves out the dividends paid since. It is set beside the whole US stock market with dividends over the same days (the Fama-French market return, from the Kenneth R. French data library), and beside the S&P 500 index, which is price only. A stock that stopped trading is counted to its last trade.
• The day pages name companies as they were called that day, and rank movers among the 500 largest by market value that day. "The day you were born" and "this day in history" use the S&P 500 index, price only, from 1928.

The Market Lab

• A condition is true on scattered days. Days within 30 calendar days of each other form one episode, so one selloff is not counted twenty times. Outcomes are measured from the first day of each episode.
• Normal is the same move from any trading day in the same years, so a result is read against what those years did anyway.
• The range beside "how often it rose" is a Wilson 95% interval. Ten episodes is a small sample; test enough situations and one will look good by luck.
• Backtests buy at the close of the day a situation starts, hold the chosen number of days (a new start inside the hold extends it), pay the cost on every buy and sell, and earn the cash rate in between. Buy-and-hold is invested on the first day and never trades.
• Days most like today compares nine measures, each scaled by how much it moved up to the day asked about (nothing after it is used), and keeps the ten closest days at least 182 days apart, leaving out the year before the day asked.

The studies in our social posts

The 18 launch findings are fixed studies run in September 2026, using index data through September 24 and stock data through September 23. They are not live signals. Unless a post says calendar year, its one-year outcome is 252 trading days; index returns exclude dividends. Overlapping dates and stocks are not independent observations.

• VIX at 40 or above: start a new episode after 63 quiet trading days. This gives ten scored episodes since 1990, with the S&P 500 higher after 252 trading days in eight; median +22.9%. The Lab's standard 30-calendar-day separation gives a different count.
• The 10-year minus 2-year Treasury yield turns negative: merge inverted stretches separated by fewer than 250 trading days. The seven merged episodes since 1976 had five positive 252-trading-day outcomes; median +13.6%. This is a different grouping from the Lab's 30-calendar-day rule.
• Insider studies: the historical sample uses public filing dates and compares officer/director purchases with quarterly common-stock observations. These are observational comparisons, not evidence that a purchase caused a return.

Patterns and your rule

• Patterns are found by a published rule on every clean stock. A hit is a move in the textbook direction of more than the 0.2% round-trip cost over 5, 21 or 63 trading days. Edge is the average move in that direction beyond the stock's own drift that year, after the cost.
• Only days priced at $1 or more with $1M average daily dollar volume (or, where volume is unknown, a $25M market value) count, and never a series flagged for bad data.
• Signals on one stock overlap in time, and stocks move together, so the true sample is smaller than the count shown. Small edges are easy to lose to real trading costs.

The game

• Orders fill at the next trading day's close; you do not know that price when you place the order. Shorts pay a borrow fee each day that rises for stocks under $5 or that have fallen hard; the 2008 short-sale ban on financials is in the game.
• The market line is what $10,000 in the whole market would have done over the same days. Ranked results are replayed on the server from the list of trades; a client number that disagrees is flagged.

Company filings and insider trades

• Source: the U.S. Securities and Exchange Commission's EDGAR system (sec.gov): the index of every filing since 1993, the company facts from each 10-K and 10-Q filed in XBRL (from 2009), and the Form 3, 4 and 5 data sets on insiders (from 2006). Source: SEC EDGAR. The SEC does not endorse this site.
• Which company is which stock: the SEC numbers each company (its CIK). A stock is tied to a CIK where our records carry it and the stock traded while that company filed; a fund or trust never is. Where no stock of ours carries a CIK (Google Inc. before Alphabet), the ticker its insiders' filings name ties it, when one of our stocks held that ticker on nine in ten of those days.
• Insider trades are open-market buys and sells (Form 4 codes P and S) by a company's officers, directors and 10% owners; grants, option exercises and gifts are left out, and so are amended filings, so nothing counts twice. Each trade counts from the day after it was filed: the first close at which anyone could have acted on it. A price more than five times away from that day's close is taken as a typing error in the filing, and the trade is shown without a dollar value.
• The Lab's insider measures add up the dollars made public in the last 91 days. They are unknown, not zero, on days the stock is not tied to its company or the SEC's data does not reach.

Bank failures

• Source: the Federal Deposit Insurance Corporation's BankFind Suite: its list of failed and assisted banks since 1934, its institution directory (state, the day a bank opened) and the quarterly reports banks file with it (from 1984). Source values and identifiers are kept as the FDIC gives them; the layout and wording are ours. No FDIC endorsement is implied.
• Assets and deposits at failure are from the last report the bank filed before it failed, as the FDIC lists them. Every dollar figure is in dollars of its day, not adjusted for inflation.
• Estimated loss is the FDIC's estimate of what the insurance fund paid out and will not recover. The FDIC says these figures are not complete for its own cases before 1986 or for cases insured by the FSLIC from 1934 to 1988; where its list has none, the page says so. A loss of zero, or below zero (the fund got back more than it paid), is what the list records.
• Assisted means the list records an assistance transaction instead of a closing: money, a guarantee against losses or help with a merger, from the FDIC or, for most savings institutions before 1989, the FSLIC.
• Assets over time come from the quarterly reports, which start in 1984: a bank's "largest" is its largest in those reports. The state is shown only for a bank that closed when it failed; the FDIC's directory keeps a bank's last address.
• A bank's name is the one on the FDIC's failure list, with its shortened words written out (NAT BK & TR as National Bank & Trust). The FDIC's directory can carry a later name for the same charter.

Report a data issue

Send the page address, the symbol or question, and what you expected to [email protected]. Corrections are noted on the page they change.