Software engineer in quant finance, distributed systems, HPC, and ML.
Boosting
OCT 14 2026 · 9 min
Boosting builds a strong model from many weak ones trained one after another, where each new model concentrates on the cases the previous ones got wrong; bagging averages independent models to cut variance, boosting adds up dependent ones to cut bias.
Random forests
OCT 12 2026 · 6 min
A random forest is bagging with one extra rule: at every split, each tree may only choose from a random handful of the features; that stops all the trees from asking the same first question, makes them disagree in more useful ways, and makes their average more accurate.
Portfolio optimisation
OCT 9 2026 · 7 min
Portfolio optimisation chooses how much to put in each asset so that, for the risk you accept, the expected return is as high as possible; it works because correlations let a mix beat its parts, and it fails when its inputs, especially expected returns, are guessed badly.
Bagging
OCT 6 2026 · 6 min
Bagging trains many copies of the same model, each on a slightly different random resample of the data, and averages their answers; the copies make different mistakes, so averaging cancels much of the noise, which makes an unstable model like a deep decision tree far more reliable.
Sharpe, Sortino and Information ratio
OCT 4 2026 · 9 min
A return on its own means little until you know the risk taken to get it; these three ratios divide extra return by risk, and they differ only in what they count as extra and what they count as risk: total swings (Sharpe), only the bad swings (Sortino), or straying from a benchmark (Information ratio).
Bias and variance
SEP 30 2026 · 7 min
A model's error on new data comes from three places: bias (it is too simple to capture the real pattern), variance (it changes too much depending on which data it happened to see), and noise (randomness no model can predict); making a model more complex lowers bias but raises variance, so the best model sits in between.
Covariance, correlation and portfolio variance
SEP 27 2026 · 8 min
Covariance and correlation measure whether two things tend to move together; in a portfolio, that one number decides how much of the risk cancels out, which is why two risky assets held together can be safer than either one alone.
Decision trees
SEP 22 2026 · 7 min
A decision tree makes a prediction by asking a short series of yes/no questions about the input, like a flowchart an expert would draw; it learns which questions to ask, and in what order, by picking at each step the question that best separates the answers.
Mean and standard deviation
SEP 19 2026 · 7 min
The mean tells you where the middle of your data is, and the standard deviation tells you how far a typical value strays from that middle; in finance the first is the return you expect and the second is the risk you take to get it.
Linear regression
SEP 14 2026 · 9 min
Linear regression draws the straight line that best fits your data, where "best" means the total of the squared misses is as small as possible; each number it learns tells you how much the answer moves when one input goes up by one.
Expected value
SEP 11 2026 · 7 min
Expected value is the average result you would get if you could repeat a risky situation many times: multiply each possible outcome by its probability and add them up. It tells you the long-run cost or gain, but not how bad one single time can be.
Regression vs classification
SEP 8 2026 · 6 min
Most machine learning answers one of two questions: "how much?" (regression, the answer is a number) or "which one?" (classification, the answer is a category), and knowing which one you are asking decides the model, the error measure and what a good result looks like.
Probability basics
SEP 5 2026 · 8 min
A probability is a number between 0 and 1 that says how often something happens in the long run, and five rules (not, or, and, given, and Bayes) let you combine them without fooling yourself.
What is machine learning?
SEP 1 2026 · 5 min
Machine learning is teaching a computer by example instead of by instruction: you show it many past cases with the right answers, it works out the pattern on its own, and then it uses that pattern on cases it has never seen.
Letting my homelab do things on its own
AUG 29 2026 · 5 min
A chat window only answers when I ask. I wanted my homelab to also do small jobs for me on its own, so I tried two AI agents, OpenClaw and Hermes Agent. How they work, how I set them up, and the surprise that most of my jobs didn't need AI at all.
Switching to a smaller model
AUG 20 2026 · 4 min
I expected a new, smaller model to be slower on my homelab, and at first it was. Then I found one setting that more than doubled its speed. In the end I switched to it, because it freed up half my memory and can read images.
Setting up my homelab
AUG 16 2026 · 5 min
I wanted my own private AI, running at home, where nothing I ask leaves the house. This is the machine I bought, how I set it up, what I tried, how local models felt to use, and what I want to try next.