> Fitting a line to a log-(log-square-root) graph using linear regression sounds awfully fishy to me. Would anyone with a better grasp of the statistics care to

> Fitting a line to a log-(log-square-root) graph using linear regression sounds awfully fishy to me. Would anyone with a better grasp of the statistics care to chime in? It's fishy, but depressingly common in Economics. There's an old joke that 'everything on a log-log plot is linear when drawn with a fat magic marker'. The errors at the top-right of the chart are much larger than the errors on the bottom left: they just don't look that way because of the log scale. They're also systematically biased at the top-left. If you want to run OLS on a timeseries, you need ergodicity (i.e. the time series forgets its own history) for the regression to be valid. The number of observations also needs to be across a time much larger than the relaxation time of the timeseries. I don't see that here. The entire thing triggers my spider sense for cranks. It's written in word and uses excel for its plots, the author runs some sort of crypto fund, and the fact that it uses 'square root of time' as a regressor is just bizzarre.