Can an AI-fitted line predict the future? Audit slope, residuals and extrapolation

Define the variables before supplying the rows

In this invented editorial exercise, x counts sorting sessions and y counts completed cards. The points are(1,2),(2,4),(3,4). They are calculation material, not a productivity experiment or evidence that more time necessarily causes a particular outcome.

The Google linear regression course explains the intercept-and-slope representation. Here the method is ordinary, equally weighted least squares with an intercept, not a line forced through zero or a model with extra features. No model was called and no business system was trained.

Derive slope and intercept explicitly

The means are x̄=2 and ȳ=10/3. The sum of centered products is(-1)×(-4/3)+0×(2/3)+1×(2/3)=2. The centered squared-x sum is 1+0+1=2. Thus slope b=2/2=1 and intercept a=10/3-1×2=4/3.

The fitted equation is ŷ=x+4/3. Simply joining the first and last point does not establish the same fit. Least squares minimizes the squared residual sum for these points; it does not force the line through every observation.

Name the error quantity and its denominator

SSE=1/9+4/9+1/9=2/3. Defining training MSE as SSE divided by the three points gives 2/9. That is not a variance estimate divided by residual degrees of freedom. A different denominator can explain another reported quantity; the number alone is insufficient.

The residuals sum to zero here, which is a useful arithmetic cross-check. It does not mean each prediction error is zero. For a related denominator distinction, see sample versus population standard deviation.

Request work that can be checked independently

Readers can use this prompt for their own exercise. The tables above are a human reference, not output measured from a named model; the illustrative error is not a product review.

Use only these invented points:(1,2),(2,4),(3,4). Fit ordinary unweighted least squares with an intercept. Show means, centered sums, slope, intercept, each fitted value, residual defined as observed minus fitted, SSE and training MSE defined as SSE/n. Keep exact fractions. Evaluate x=2.5 and x=4, labeling interpolation versus extrapolation. Do not invent an observed fourth point, a confidence level, or a causal effect.

Attach conditions to 23/6 and 16/3

At x=2.5, the line returns 23/6≈3.833 within the observed x range: interpolation. At x=4, it returns 16/3≈5.333 outside 1–3: extrapolation. Both are model values, not counts of cards already completed.

Cards are counts while a line can return fractions. A value may describe an average relationship, but cannot silently become five completed cards and a third of another. More distant inputs or a changed sorting process may follow a different relationship. Decisions need relevant data and independent validation; these three rows supply no confidence interval or future accuracy rate.

Repair the layer that actually failed

Wrong parameters require recomputing means and centered sums. Correct parameters labeled as actual x=4 outcomes require repairing observation versus prediction. A statement that time caused the output change needs study-design evidence. Use correlation, confounding and experiment design to examine that separate gap.

Keep the three original rows, fitting convention and reference arithmetic. A small exercise can expose computational and interpretive errors without certifying a production forecasting system for every task.