Data science can feel like a recent invention, something that emerged from Silicon Valley in the 2010s alongside big tech and big data. In reality, the foundations stretch back nearly two centuries, built up gradually by statisticians, mathematicians, and computer scientists long before the term itself existed.
19th century: the first models
The story starts in the early 1800s, when astronomers needed a principled way to fit observations to predictions. The method of least squares, developed independently by Gauss and Legendre, gave them exactly that: a technique for finding the line that minimises the total error across a set of data points. It is the earliest form of linear regression, and it remains one of the most widely used tools in statistics today.
1930s to 1970s: classification and generalisation
The next major wave came in the mid-20th century, when statisticians turned their attention from prediction to classification. Ronald Fisher developed linear discriminant analysis in the 1930s to separate groups based on measured features. Logistic regression followed as an alternative approach for modelling binary outcomes.
These were powerful but specialised tools. In the early 1970s, statisticians unified them under a single theoretical framework: generalised linear models. GLMs extended the logic of linear regression to handle a much wider range of response types, from proportions to counts, and provided a coherent foundation for the field.
1980s: computation opens new doors
By the late 1970s, the statistical toolkit was rich but largely limited to linear relationships. Non-linear modelling was theoretically possible but computationally prohibitive on the hardware of the time. As processing power improved through the 1980s, that barrier fell.
Two important methods emerged as a result. Classification and regression trees offered an intuitive, interpretable way to model non-linear relationships by recursively splitting data on the most informative features. Generalised additive models provided a more flexible extension of GLMs, allowing smooth, non-linear terms for individual predictors while retaining much of the interpretability of linear models.
The same decade also saw neural networks gain serious attention. Inspired loosely by the structure of biological neurons, multi-layer networks trained with backpropagation showed that machines could learn complex patterns from data, not just fit pre-specified equations.
1990s: support vector machines and the modern era
Support vector machines arrived in the 1990s with a different geometric intuition: find the boundary between classes that maximises the margin between them. With the kernel trick, SVMs could also handle non-linear decision boundaries without explicitly mapping data into higher-dimensional space. For a period, they were state of the art across many classification tasks.
The democratisation of the field
What changed most in the decades that followed was not the invention of entirely new methods, but access. Tools like Python, with libraries such as scikit-learn, NumPy, and pandas, took techniques that had previously required specialist mathematical training and made them available to anyone willing to learn. A student could implement a random forest or a neural network in an afternoon. The barrier to entry dropped dramatically, and the field grew accordingly.
That accessibility is what drew me in. The history of data science is a history of ideas that started in astronomy and statistics, were shaped by decades of computational progress, and eventually became tools that almost anyone can pick up and apply.