Posts

Showing posts from November, 2026

The VIX against what the market did next

Five days out of six. Over the last ten years, that's how often the VIX stood above the volatility the S&P 500 then delivered. Second of three volatility posts here. Last Friday's post worked out historical volatility from ten prices, a number that looks backwards. This one puts it beside the number the options market looks forward with. Cboe calculates the VIX from S&P 500 option prices and quotes it as an annualised percentage for the coming 30 calendar days. Thirty calendar days is about 21 trading days, so every reading can be marked against what followed. I did that for 2,468 trading days, December 2016 to October 2026, with the index closes the St Louis Fed publishes. The VIX averaged 18.7. Realised volatility over the following month averaged 15.1. On 84 per cent of days the VIX was at or above the outcome, usually by around 4.6 points, and in 2023 it was above on every day. Why pay for it, then? Because when it's wrong, it's wrong in one direction and by...

Historical volatility with a pencil and ten prices

Ten closing prices, starting at £100 and finishing at £111. That's all the one-page volatility guide works with, and it's enough to get to an annualised figure of 24.5 per cent without opening a spreadsheet. It's the first of three posts on volatility here, one a week, following the video series of the same name. The prices run 100, 102, 101, 103, 105, 107, 110, 108, 109 and 111. Each neighbouring pair gives a log return, the natural log of today's price over yesterday's. Nine of them: 1.98, −0.99, 1.96, 1.92, 1.89, 2.77, −1.83, 0.92 and 1.82 per cent. They average 1.16 per cent a day. Far too high for anything real, and on purpose: a series that climbs keeps the arithmetic tidy. With a year of real closes the daily mean comes out close to zero. Then the spread around that average. Take the mean off each return, square what's left, add the nine squares and divide by eight, not nine. The square root of that is 1.54 per cent a day. Dividing by eight is the usual s...

The crack spread, cut down to one page

Image
$47.05 a barrel. That's what the 3-2-1 crack spread came to on the closing prices of the first week of November 2026, and it's the number at the top of a one-page guide I've just put up. Why one page? The full article on refinery margins covers hedging, the trouble with crude quality and the trouble with location. Someone in their first week on an energy desk wants less than that before lunch. The formula, one worked example and the seasonal shape will do. The sum is short. Gasoline (RBOB) closed the week at $2.7227 a gallon, diesel (ULSD) at $4.1983 and crude (WTI) at $87.96 a barrel. Multiply the product prices by 42 to turn gallons into barrels: two barrels of gasoline come to $228.71 and one of diesel to $176.33. Take away three barrels of crude, $263.88, and you're left with $141.16. Divide by three for the margin on each barrel of crude. For scale: across five years of weekly prices, starting November 2021, the average was $32.80. Its low was $14.60, in September ...

Where the code from my articles lives now

Most of my data articles end with a short Python script. Until this week, the only way to run one was to copy it out of the web page, paste it into an editor and hope the indentation survived. That has changed in two ways. On Kaggle, five of those scripts now run as notebooks. You open one, press run, and it prints the figures from the article. Nothing to install. The five cover position sizing and drawdowns, plastics resin output as a lead on manufacturing, implied volatility from Black-Scholes, 25 years of market correlations, and how far non-farm payrolls get revised. On OSF, the Open Science Framework, fourteen scripts sit as plain files in one project, sorted into folders: volatility, correlation and spreads, energy, macro and risk. They carry the MIT licence, so anyone can reuse them. OSF also gives the project a permanent address, which matters if someone wants to cite the method rather than a web page that might move. Why both? Kaggle is for the reader who wants to see the numb...