where
A repository of adhoc analyses and algorithmic demonstrators using various programming languages.
I'm a fan of proof oriented languages like Prolog and Haskell because I think they are the closest thing we have to a "language for thought". Even so, it is uncommon to see analyses in these languages outside of finance, so I was compelled to produce some examples for posterity.
I'm also a fan of more specialised programming languages like Fortran, which I had just learned at the time of writing to give the 1BRC challenge a go. I was amazed by how easy it is to pick up and produce high performance code with, so I'm very excited to see what uses I can find for it.
The most buoyant turn-key eco-systems for ML and statistics are centred on Python, R and Julia. Outside of these, the common complaint is library support, so part of what I aim to show is that if we pay attention to what we are doing when writing algorithms and performing analyses, there is often not much we need that we cannot quickly write yourselves.
UK property model -- A model to
interpolate the price series of a UK property from
public price paid data by setting up a large sparse
regression. Two solutions are given: an approximate
solution in Haskell, and an exact solution in Fortran.
haskell, fortran, regression, sparse,
parameter-estimation, autodiff, optimisation,
lapack
Jointly learning transformations and fitting a regression
-- A 5 parameter non-linear pipeline composed of a two
Yeo-Johnson transforms, a sigmoid function (which bisects
the model), and a linear regression. It is jointly fitted
using "threshold accepting" annealing (implemented
herein). haskell, feature-learning, regression,
global-optimisation, parameter-estimation.
Convolutional logistic regression -- Next up!
Classification of digits from the
MNIST 1D dataset
by jointly fitting convolutional filters and a logistic
regression model using the thresholding accepting
algorithm. fortran, feature-learning,
classification, autodiff