Read a fitted model before using its estimate

Give each predictor its own coefficient

A two-predictor additive model has the form y = b + c1*x1 + c2*x2. Each coefficient describes a fitted change while the other predictor stays fixed. Estimating two separate one-variable slopes generally produces a different model.

For an illustrative model y = 4 + 2*x1 − 3*x2, the input x1=2 and x2=1 gives 5. Leaving out the intercept gives 1; swapping the coefficients gives 0. Map each input to its actual coefficient before calculating the estimate.

Match the returned order to the source columns

When B contains X1 and C contains X2, LINEST returns the X2 coefficient, the X1 coefficient and then the intercept. The two-predictor task selects each output with INDEX rather than assuming the leftmost returned number belongs to B.

Keep each measured Y paired with both predictors from the same row. A joint row permutation preserves the fit. Sorting one source column independently creates new observations and changes the relationship being fitted.

Measure residuals on the fitted scale

An additive residual is measured Y minus fitted Y. SSE adds the squared residuals. With eight observations, two independent predictors and a fitted intercept, residual standard error is SQRT(SSE/5). R-squared compares SSE with the total squared deviations around the measured mean.

For a multiplicative model y = b*m^x, fit a line to log(Y). LOGEST returns the multiplier m and level b. Its residual standard error and R-squared describe log(Y); a raw Y difference uses another scale. All observed Y values must be positive.

Keep period units and the new input explicit

A multiplier of 0.8 gives a fractional change of −0.2 for one period unit. Two units multiply the fitted level by 0.8 squared. Irregular period gaps therefore belong in known X; row numbers cannot replace them.

TREND and GROWTH calculate a model estimate at the supplied new inputs. Training fit does not measure future error. Check held-out observations before treating an estimate as a forecast, and examine data collection before making a causal claim. Collinear predictors also require a model decision because separate coefficients are no longer uniquely identified.

Try the idea

• Fit a two-predictor model →
• Fit multiplicative change →
• Regression model course →
• Explore single-predictor plots →

References

• Microsoft LINEST ↗
• Microsoft LOGEST ↗
• Microsoft TREND ↗
• Microsoft GROWTH ↗

Examples and explanations are original. Source links describe the related Excel features; native Excel workbook testing is separate from browser learning checks.