Logistic Regression Predictions With Sigmoid Intuition

Logistic regression models P(y = 1 | x), a probability in [0, 1], by passing a linear predictor through the sigmoid. The model is logit(p) = ln(p / (1 − p)) = β₀ + β₁x₁ + ⋯ + βₖxₖ, which inverts to p = 1 / (1 + e^−(β₀ + β·x)). The output is a probability, not a class label. Classification comes from picking a threshold (default 0.5) and predicting 1 when p ≥ threshold.

The interpretation trap: a coefficient β = 0.5 does not mean "probability rises by 0.5 per unit of x." It means log-odds rise by 0.5, equivalent to multiplying the odds by e^0.5 ≈ 1.65. The probability change depends on where you sit on the sigmoid: small near the tails, largest near p = 0.5. For inference, report odds ratios e^β. For prediction, report p evaluated at concrete x. And for imbalanced classes, 0.5 is rarely the right threshold. Pick it from the ROC curve based on the cost ratio of false positives to false negatives in your application.

Logistic regression first computes a linear predictor z = β₀ + β₁x₁ + β₂x₂ + … where βs are coefficients and xs are feature values. This z can range from negative infinity to positive infinity—it's the log-odds of the positive class.

The sigmoid function σ(z) = 1 / (1 + e^(−z)) transforms z into a probability between 0 and 1. Large positive z yields probability near 1; large negative z yields probability near 0; z = 0 yields exactly 0.5. The S-shaped curve is steepest at the midpoint—small changes in z produce the largest probability shifts when probability hovers around 0.5.

Understanding the sigmoid matters because it shows why coefficient effects are non-linear in probability terms. A one-unit increase in a feature always adds the same amount to z (the linear predictor) but adds a variable amount to probability depending on where you sit on the curve.

Provide the intercept (β₀) and coefficients for each feature. In real applications, these come from training a model on labeled data via maximum likelihood estimation. Here, you enter them manually to explore how changes affect predicted probabilities.

Then input feature values (x₁, x₂, …). The tool computes z, applies the sigmoid, and reports the probability of class 1. The probability of class 0 is simply 1 − p.

Experiment by varying feature values while holding coefficients fixed. Watch how probability responds—a large positive coefficient on a feature makes probability climb steeply as that feature increases; a negative coefficient makes probability fall.

A classification threshold converts probability into a binary decision. The default 0.5 means "classify as positive if p ≥ 0.5." But 0.5 isn't always optimal—it depends on the cost of errors.

Lower the threshold (e.g., 0.3) when missing positives is expensive. Medical screening for a serious disease might use a low threshold to catch more true cases, accepting more false alarms. Raise the threshold (e.g., 0.7) when false positives are costly—spam filtering on critical inboxes, for instance, where mislabeling a legitimate email hurts more than letting some spam through.

Threshold selection trades off precision (of those labeled positive, how many truly are) against recall (of all true positives, how many did we catch). Moving the threshold doesn't change the underlying probability estimate—only the decision boundary.

Odds are p / (1 − p). If probability of success is 0.8, odds are 0.8 / 0.2 = 4 (four-to-one in favor). The coefficient β in logistic regression tells you how much the log-odds change for a one-unit increase in x.

The odds ratio OR = e^β translates that to a multiplicative effect on odds. If β = 0.5, OR ≈ 1.65, meaning odds increase by 65% per unit increase in x. If β = −0.3, OR ≈ 0.74, meaning odds multiply by 0.74 (a 26% decrease).

Odds ratios are often easier to communicate than raw coefficients. "Each additional year of experience multiplies the odds of promotion by 1.3" is more intuitive than "the coefficient on experience is 0.26."

This calculator demonstrates how logistic regression makes predictions given coefficients you supply. It does not estimate coefficients from data—that requires maximum likelihood estimation on a labeled dataset, typically done in Python (scikit-learn), R, or specialized software.

Real models also need validation: train/test splits, cross-validation, calibration checks, and diagnostic plots (ROC, confusion matrices). Coefficients have standard errors and confidence intervals that this tool doesn't compute.

Treat outputs here as educational explorations, not production predictions. For any consequential decision—loan approvals, medical diagnoses, hiring—use validated models built by qualified professionals with proper data governance.