Original Research: The Datasets Behind Every Figure

Michael Whitman · Trading since 2017, full-time since 2020 · Published 2026-09-22 · Updated 2026-09-23

Most figures on finance sites are copied from other finance sites. The figures here are measured, and this page lists the files they are measured from, so any number on any page can be checked against its source rather than taken on trust.

How it works

Each dataset is a raw file plus the measurements taken from it. A page names its file in its sources, and a figure goes on a page only if it can be recomputed from that file. Before this page was published, every headline number below was recomputed from the raw series and checked against the stored summaries, and all of them matched.

The video study

The study collected YouTube search results for 1,409 trading and finance queries in August 2026. Those searches returned 31,760 results. Many videos rank for more than one query, so duplicates were removed by video ID, which leaves 24,971 unique videos from 10,608 channels. Each row holds the title, length, view count, channel and the query that found it.

The per-term counts that pages quote (“132 videos with index funds in the title”) come from matching a term against those titles. The raw rows are not published, because they carry other creators’ titles and description text; the summary, the per-term counts and the per-broker counts are.

The S&P 500 series

Daily closing levels of the S&P 500 from January 3, 1950 to September 18, 2026: 19,300 closes. The series is the index level, price only. Dividends are not in it, which matters for any long-horizon figure and is covered in the limits below.

Three files are measured from it. The drawdown summary tracks the running record close and how far below it each day sat. The bear-market list marks every record close, finds the lowest close before the next record, and keeps the episodes where that fall reached 20%. The holding-period table asks, for every start date, whether the index closed higher a fixed number of trading days later, with one year counted as 252 trading days.

Bitcoin and gold

Month-end closes in US dollars: 144 months of bitcoin from September 2014 to August 2026, and 313 months of gold futures from August 2000 to August 2026. Both come from the same public price provider as the index series, and both are published here as they were measured, one row per month.

A worked example

Take the largest number on the site and redo it. The S&P 500 made a record close of 1,565.15 on October 9, 2007. The lowest close before it made another record was 676.53 on March 9, 2009.

676.53 / 1,565.15 = 0.432. Subtract that from one and the fall is 56.8%. Open the daily closes file, find those two rows, and the division gives the same answer. The next close above 1,565.15 came on March 28, 2013, five and a half years after the peak.

The holding-period figures work the same way. There are 19,048 start dates with a close 252 trading days later, and 74.6% of them ended higher. That is not a forecast. It is a count of what already happened, and anyone with the file can repeat the count.

The original data

The video study. Of 24,924 videos with a visible view count, the median has 10,767 views; counted the way the topic pages count, with the 47 missing counts as zero, it is 10,684. 49.0% have fewer than 10,000 views and 3.2% have more than a million. Together they hold 4,061,826,125 views. The median length is 12.1 minutes, and only 5,827 of the 24,971 videos had description text in the search results, which limits what can be said about what the videos contain.

The S&P 500, 1950 to 2026. The index closed up on 53.1% of days, and 59.1% of ten-day windows ended higher. It sat below a previous record close on 92.0% of trading days. The longest stretch without a new record ran from January 11, 1973 to July 17, 1980, about 7.5 years.

Bear markets. There were 11 falls of 20% or more from a record close. The smallest was 21.5% (1956 to 1957), the median 33.5%, and the largest 56.8%. From peak to the next record took a median of 2.0 years, as little as half a year (2020) and as long as 7.5 years (1973 to 1980).

Holding periods. The share of windows that ended higher rises with time: 53.1% at one day, 61.8% at one month, 74.6% at one year, 84.2% at five years and 93.1% at ten. Every one of the 14,260 twenty-year windows ended higher, and the worst of them was still up 53.8%, before dividends.

Bitcoin. From $386.94 at the end of September 2014 to $78,548.63 at the end of August 2026. The highest month-end close was $115,758.20 in July 2025. The worst fall measured on month-end closes was 75.6%, from the December 2017 peak to January 2019. 55.9% of months closed up; the best month gained 69.6% and the worst lost 37.8%.

Gold. From $278.30 in August 2000 to $4,481.50 in August 2026, 16.1 times the starting price, or about 11.3% a year. Its worst month-end fall was 42.0%, from August 2011 to December 2015, more than four years of losing ground. 56.4% of months closed up.

When it fails

Price-only index figures understate what a holder earned. The S&P 500 series leaves out dividends, which over decades are a large share of the total return. A “worst 20-year window” of +53.8% is the floor on price alone; with dividends reinvested it would have been higher. Read every long-horizon figure here as the conservative version.

Month-end closes hide the path inside the month. Bitcoin’s 75.6% fall is measured from one month-end to another. Intraday and intra-month lows went further, so the real worst case for anyone who held through it was deeper than the monthly file can show. The same applies, less violently, to gold.

Gold here is a futures price, not the spot price. Futures and spot usually sit close together, but they are not the same series, and a fund that holds bullion charges a fee the futures price does not.

The video study is a snapshot of search results, not of YouTube. It contains what search returned for these queries in August 2026. Views keep growing after collection, rankings change, and a video that never ranked for any of the 1,409 queries is not in it. A title that names a term is not proof the video teaches that term well, which is why pages quote title counts and view figures and never claim to know what a video says.

None of it predicts anything. A share of past windows that ended higher is a description of the past. The next twenty years can be the first one that breaks the pattern, and every page that uses these files says so where it matters.

License and reuse

The measurements, summaries and counts on this page are free to reuse with attribution to Whitman Trading and a link to this page (Creative Commons Attribution 4.0). The raw price series are redistributed so the figures can be checked; the prices themselves come from the providers named above, and their terms apply to any reuse of the raw prices.

Related

The rules every page follows, including the recompute rule this page is built on, are set out in how this site is made. For what a fall from a record close means for a single position rather than an index, the drawdown page works through the arithmetic of recovering from one. The eleven episodes above are discussed as a group on the bear market page, and the long-horizon side of the same series is where the index fund page starts. The monthly bitcoin file underpins the bitcoin page, and the gold file does the same for gold.

This page is educational, not financial advice. Test every idea on your own charts before risking money.