Regression in Business: Prediction, Persuasion and False Confidence

Regression is useful because it turns a pattern into a measurable relationship. It becomes dangerous when a fitted relationship is treated as a causal law, a high in-sample fit substitutes for validation, or narrative claims are copied without checking the workbook calculations.

Reference report · No directional market call

This report combines the regression primer and presentation with an anonymised retail-research teaching workbook. The workbook contains 35 customer observations and seven variables. Its numerical outputs are retained as educational data, not as findings about an identifiable retailer or a verified customer population.

Define the prediction before fitting a line

The outcome in the workbook is monthly purchase spending. Potential predictors include monthly income, age, time on the website, daily advertisements seen and satisfaction score. The simple regression uses income alone. Monetary values are measured in thousands of rupees, so a unit change in the income variable means ₹1,000.

A simple linear model writes:

$$Y_i=\beta_0+\beta_1X_i+\varepsilon_i.$$

Here $Y_i$ is monthly spending for observation $i$, $X_i$ is monthly income, $\beta_0$ is the population intercept, $\beta_1$ is the population slope, and $\varepsilon_i$ captures variation not represented by the line. Estimated coefficients are denoted $b_0$ and $b_1$; a fitted value is $\hat{Y}_i=b_0+b_1X_i$.

The regression question is whether the observed relationship can help explain or predict spending within the relevant population and range. It is not whether raising a customer’s income would necessarily cause spending at this retailer to rise by the estimated amount.

What ordinary least squares does

Ordinary least squares chooses the coefficients to minimise the sum of squared residuals. A residual is the observed spending minus fitted spending. Squaring makes positive and negative errors contribute positively and gives larger errors more weight.

For a simple regression with an intercept:

$$b_1=\frac{\sum_i(X_i-\bar{X})(Y_i-\bar{Y})}{\sum_i(X_i-\bar{X})^2}, \qquad b_0=\bar{Y}-b_1\bar{X}.$$

In words: Estimate the slope from how income and spending move together, then set the intercept so the line passes through their sample averages.

The bars denote sample means. The numerator measures co-movement; the denominator measures variation in income. If every sampled income were identical, the slope would not be identifiable from these observations.

A scatter plot should precede interpretation. It can reveal a curved relationship, clusters, outliers and changes in spread. The best-fitting straight line can still be a poor description when the relationship is nonlinear or when a few observations dominate the fit.

The retained numerical result

The workbook’s live output reports the following:

Workbook results and their interpretation

The fitted equation is therefore $\hat{Y}=-5.721732+0.305877X$, with both monetary variables in thousands of rupees. At income ₹55,000, the fitted monthly purchase is approximately ₹11,101.50.

That is a fitted conditional mean under the model. It is not a guarantee for an individual customer. An individual prediction interval includes variation around the mean relationship and is wider than a confidence interval for the mean response at the same income.

Keep narrative and calculated results consistent

Narrative examples that differ from the live workbook results

Those values are not the live outputs listed above. They are retained here as a documented discrepancy in the teaching material, rather than silently substituted for the computed result.

An unusually high R-squared in a small educational dataset should prompt caution. It can make calculation mechanics easy to see, but it does not establish that a similar relationship will hold in commercial customer data. Before deployment, review how records were generated, whether any outcome information entered the predictor, and whether the relationship survives a genuinely separate sample.

The rule for reporting is simple: every narrative number should reconcile to the exact analysis being described. A template explanation must not survive unchanged when the underlying inputs have changed.

Read fit, uncertainty and causality separately

R-squared measures the share of sample variation in the outcome represented by the fitted model relative to an intercept-only benchmark. With an intercept it can be written as:

$$R^2=1-\frac{SSE}{SST},$$

where $SSE$ is the sum of squared residuals and $SST$ the total squared deviation of observations from their sample mean. The workbook gives $SST=1050.991429$ and model-explained sum of squares of approximately 1041.480471, in squared thousands of rupees.

A high R-squared does not establish causation, guarantee accurate future predictions or show that all important predictors have been included. A low R-squared does not automatically make a coefficient uninformative. Fit and inferential uncertainty answer different questions.

The slope t-statistic divides the estimated slope, less its hypothesised value, by its standard error. The source reports an F-statistic of approximately 3613.606 and a very small model p-value, about $2.66\times10^{-35}$. In a simple regression with one slope and the usual test, the F-statistic equals the squared t-statistic. This numerical consistency does not validate the sampling design or the economic interpretation.

The benchmark test asks another question

The same workbook compares average spending with a ₹8,000 benchmark. It assumes a known population standard deviation of ₹3,500.

Benchmark-test outputs under the supplied standard-deviation assumption

The known standard deviation is a supplied teaching assumption. The workbook describes it as coming from industry research, but the publication does not have an independently verified underlying population estimate establishing its applicability to these 35 observations. A real analysis must test that assumption rather than inherit it from a label.

The benchmark comparison and the income regression are not interchangeable. The first asks whether a population mean differs from a specified benchmark under sampling assumptions. The second estimates an association across sampled customers. A selected sample of higher-income, satisfied customers can produce a high mean without establishing that the retailer’s overall customer population outspends the benchmark.

Check assumptions at the level they matter

Different assumptions support different claims. Linearity concerns the conditional mean specification. Exogeneity concerns whether omitted factors in the error are systematically related to included predictors. Independence or correctly modelled dependence matters for uncertainty. Constant variance supports familiar standard-error formulas; appropriate robust methods may be needed when it fails.

Normality of errors is useful for exact small-sample inference under the standard model. It is not a prerequisite for calculating least-squares coefficients, and a sample size above thirty does not automatically resolve every distributional problem.

Regression diagnostics and practical checks

Satisfaction is particularly revealing. In the workbook it is clustered between 5 and 9. The absence of low ratings may reflect the teaching construction or a selected customer sample. It should not be interpreted as proof that dissatisfied customers do not exist.

Avoid extrapolation and false “all else equal” language

The supplied income range is ₹27,000–₹90,000. The negative intercept at zero income is outside that observed range and has little practical meaning. Extrapolating the line far above the range can also produce implausible spending predictions because marginal spending propensities may change.

A single-predictor regression does not hold other measured predictors constant. Its slope describes the simple association in that specification. “All else equal” requires a carefully defined conditional interpretation and still does not establish causal identification. Adding more predictors can improve prediction, introduce instability, absorb a causal pathway or condition on inappropriate variables; selection needs domain reasoning.

Validation should match the intended use. If predicting future months, use time-respecting validation. If serving new customer groups, assess performance across those groups. Evaluate errors in money and in the decisions they affect, rather than judging solely by a percentage fit statistic.

Distinguish group comparisons from continuous relationships

The retail workbook also compares high-income and low-income customers using a Welch test. Such a comparison answers a different question from estimating a continuous income slope. A group threshold discards some information, and selecting the threshold after inspecting spending can exaggerate the apparent contrast.

State how groups were defined, their sample sizes, means and variability. Welch’s method accommodates unequal estimated group variances under its assumptions; it does not remove selection bias or establish that income caused the difference. A meaningful analysis should also test whether the result depends heavily on a few customers close to the dividing threshold.

For operational use, compare the proposed model with a simple baseline such as predicting the sample mean or a pre-existing segment mean. A statistically impressive model may offer little incremental decision value if the baseline already makes the same choices. Evaluate the cost of large errors separately when over-predicting a customer’s value leads to expensive acquisition or service commitments.

Model monitoring then extends the original analysis. Track whether input distributions, prediction errors and segment performance change over time. A deteriorating relationship may require recalibration, a new predictor set or withdrawal from the use case. The original p-value does not grant permanent validity to the model.

Translate the analysis into a bounded decision

The educational result supports practising model interpretation and identifying what would need validation before a commercial use. It does not justify a general instruction to target only high-income customers. Such an action also depends on acquisition cost, contribution margin, service cost, retention, fairness constraints and the risk of excluding valuable segments.

A defensible report separates the fitted relationship, evidence quality and proposed action. It might recommend using income as one candidate predictor in a properly sampled validation exercise, comparing it with simpler baselines and measuring commercial value before changing targeting. That recommendation follows from the strength and limits of the evidence without pretending that a teaching dataset establishes a market fact.