Published:
I spent an afternoon on a hyperparameter sweep for this model and it made things slightly worse. I added some weather and calendar interactions, and those hurt too. The thing that actually took the model from mediocre to good was a single plot of where it went wrong, and four features that followed from it. The model itself never changed.
Published:
One of the baselines I built for this project scores 0.975 accuracy and 0.948 F1, which comfortably beats my LSTM. It is also useless as a forecaster, and the reason it is useless is the most useful thing I got out of the project.
Published:
There is a failure mode that most deployment tutorials skip past: you retrain, the new model is worse than the one you already have, and it replaces it anyway. Nothing errors. Nothing looks wrong in the logs. The predictions just get worse, and you find out later than you would like.
Published:
For image classification on a small dataset, the standard advice is to take a network pre-trained on ImageNet, freeze the backbone, and train a new classifier head. It is fast and needs almost no compute. On Galaxy10 it came last — behind full fine-tuning, which I expected, and also behind a 0.31M-parameter CNN I trained from scratch, which I did not.
Published:
Before building any model, you need to understand the data well enough to make defensible modeling decisions. This post walks through how I approached exploratory data analysis on the UK Department for Transport’s 2023 road accident dataset — 104,258 collisions and 189,815 vehicle records. The EDA directly shaped the dual-model strategy I describe in my class imbalance post.
Published:
When 76% of your labels belong to a single class and the rarest class sits at 1.4%, standard classifiers will happily predict the majority class every time and report impressive accuracy. This post walks through what I learned building a severity classifier on 104,258 UK road collisions from the Department for Transport’s 2023 STATS19 data, and why the numbers that looked great at first turned out to be completely wrong.
Published:
If you have ever pasted an entire research paper into ChatGPT or Claude and watched your token budget evaporate, you know the problem. A typical 10-page paper burns 8,000-12,000 tokens, yet the model only needs a few hundred to answer most questions about it. I built LatexForLLM to fix this. It parses LaTeX documents into a typed graph and retrieves only the sections, equations, and figures that matter. On benchmark tasks against a realistic 200-line paper, graph-based retrieval cuts word count by ~54% on average (up to ~80% for focused queries) compared to pasting the full document.
Published:
Most data work involves going back and forth between SQL databases and Python. You write a query to pull what you need, load it into a DataFrame, do your analysis, maybe write results back. Pandas has built-in support for this workflow, and once you set it up, you rarely need to leave Python to interact with your database.
Published:
In this notebook we will carry the trend analysis of technology stocks.
Published:
Code for portfolio optimization using Modern Portfolio Theory.
Published:
We will discuss the time series analysis using finance data. The techniques like Moving Average (MA) , Autoregressive (AR) and Autoregressive Integrated Moving Average Model (ARIMA) will be dicussed. For modelling the time series we will be using the statsmodel library for the data acquired using Yahoo finance api.
Published:
We create a portfolio of stocks from American markets, analyze their performance and try to acess the risk in future.