Machine Learning

I understand why Prophet gets picked up so often. You give it a date column and a value column, fit the model, ask for the future dataframe, and it gives you a chart that looks like somebody has already put it in a planning deck (I’m sure that in the last 30 days you’ve met at least 1 person who has done this). Anyways, there is a trend, there are changepoints, there are uncertainty bands. It feels much more finished than a lot of time series work feels in the beginning. ...

I have spent around seven years working in e-commerce in one way or another. I have worked with brands doing a couple hundred thousand dollars a month. I have helped smaller brands that were barely starting up and just trying to get the foundations right. I have run my own e-commerce stores and I have also worked with high-end fashion brands across five countries. Most of my work was as a technical marketer building websites, marketing workflows, funnels and helping brands sell more stuff. The work also involved looking at people who bought once and never came back, people who abandoned carts halfway through, campaigns that pulled in customers who had no reason to stay, and the general pain of trying to make a store feel worth returning to. ...

I have been spending a lot of time around AI agents lately. Agents that can read files, call tools, update things, research things, write things, and generally make you feel like the computer might be about to become a coworker. Which is exciting. Also slightly stressful if you think about what it has access to for more than 30 seconds. Then there is ARIMA. ARIMA looks like it was named by someone who wanted to make sure nobody accidentally found statistics too exciting. It does not have a chat interface, cannot browse the web and definitely cannot write a follow-up email with “just circling back” in it. It looks at a sequence of numbers and asks some basic questions. ...

My old churn analysis capstone post has somehow become one of the most popular posts on this blog over the last year. I did not expect that. At the time, churn looked like a neat data science problem to me. You get a dataset, clean it, engineer some features, train a model, check the metrics, maybe explain feature importance, and then recommend a few retention ideas. It is a good project shape. Business problem, dataset, model, metric, interpretation. Very neat, if I may call it that. Very portfolio-friendly. ...

I finished my data science and machine learning master’s a week ago. That sentence feels strange to write because its been a crazy 2 years. Classes, assignments, tests on the weekends and full on work and family life through the week. It feels great to have my weekends back after 2 years. I should probably have a cleaner feeling about it. Relief, pride, maybe a very professional LinkedIn post with a certificate photo and a stupid caption about growth. Maybe I will do just that after finishing this article. Anyways, I do feel proud, obviously. It was a lot of work. Doing a master’s while working full time is not exactly what I would call a chill hobby. ...

Why I’m Picking Up the Math Now I come from product & growth marketing and I’m doing my masters in Data Science and Machine Learning. Most of my work has been about making technical products understandable. Shaping go-to-market plans, writing positioning for AI features, and working across teams to make sure what we build actually makes sense to the people we’re building it for. I’m also on a path to becoming a full-stack machine learning engineer. That means going beyond talking about models. I want to build them. Understand them. Debug them. Know when they’re lying. ...

When Copy‑Paste Stops Working In marketing, we live and die by templates. Landing page templates. Ad copy frameworks. Campaign recipes. Machine learning wouldn’t be that different, right? Just plug in your dataset, borrow someone’s notebook from Kaggle, tweak a few parameters, and be done with it. That’s how a lot of people approach it. And to be fair, it kind of works, until it doesn’t. You get a model that runs, a prediction that looks reasonable, a chart that impresses in a meeting. But under the surface, there’s often a gap. You don’t really know what’s going on. ...

K-Means++ is a clever upgrade to K-Means that fixes its biggest flaw: random initialization. Instead of picking all k centroids at random, K-Means++: Picks the first centroid randomly from the data points. For each remaining point ( x ), compute its shortest distance ( D(x) ) to the nearest chosen centroid. Choose the next centroid from the dataset with probability proportional to ( D(x)^2 ). Repeat until ( k ) centroids are selected. This spreads centroids out more effectively and leads to: ...

If you are trying to decide between ARIMA and SARIMA, the thing you are really checking for is seasonality. ARIMA is for a series with a trend or short-term dependence, but no predictable repeating cycle. SARIMA is for when you have that same stuff plus a repeating pattern, like monthly demand going up every December or daily traffic acting differently every weekend. Pretty much. The annoying bit is that seasonality can look like random mess until you actually plot the data. Also, a seasonal model sounds more complete, so it is tempting to throw one at everything. I have done that. Forecasting gets annoying pretty fast when the model is more complicated than the data needs it to be. ...

If you read part 1 of this series, ala Churn Prediction for E-Commerce with Predictive Modelling, you know I recently wrapped up a full end-to-end churn prediction project as part of my postgrad program. That article was the 30,000-foot view – the business problem, the segmentation insights, the high-level model results. With this one I simply walk you through the code. But instead of just dumping code snippets for you to copy pasta, I want to walk you through what I actually did and, more importantly, why it matters. ...