Fitting Linear Models to Data

Real data rarely falls in a perfect line. Plant heights, test scores, prices — any two related quantities measured in the real world come with noise, small ups and downs from measurement error or natural variation. If you connected every dot, you would get a jagged zigzag that fits the noise instead of the pattern. A straight-line model ignores that noise and captures the overall trend, which is what lets you make a sensible prediction for a week, or a value, you never measured.

The line carries two numbers that mean something in context. The slope is the rate of change: how much the output goes up or down for every one unit increase in the input. The intercept is the starting value: what the model predicts when the input is zero. For a plant growing over time, the slope is the growth rate in centimeters per week, and the intercept is the height at week zero. Because the line is fit to minimize the total leftover gaps across all the points at once, not to pass through any single one, it almost always misses every plotted point while still being the best overall summary.

This same idea is really solving for the two unknowns, slope and intercept, that pin down a linear equation, which is why fitting a line connects directly to systems of two linear equations: two conditions on a line's behavior are enough to determine both its slope and intercept, just as two equations pin down a solution point.