ONLYOFFICE STEYX: Regression Residual Error with a Complete Worksheet
Rebuild ONLYOFFICE STEYX from squared residuals, understand the n-minus-two denominator, and distinguish residual spread from a prediction interval.
Verification status: See the evidence and test scope described in this guide. How formula examples are checked
Quick answer: =STEYX(B2:B6,A2:A6) has an independently expected result of approximately 0.894427191 for the data below. It summarizes spread around a fitted straight line in the y-variable’s units. It does not return the slope, correlation or a forecast interval for an individual new x.
Availability and version history
- Introduced: The first introduction release and date are not established by the checked sources.
- Availability: Currently documented in ONLYOFFICE’s rolling function help; no final supported-version endpoint is established.
- Scope: Shared Spreadsheet Editor documentation. Exact availability in individual Docs, Desktop and mobile builds is not established by this page.
- Checked: October 14, 2026, against the official STEYX reference. Examples were independently calculated and compared with the vendor screenshot; these cases were not tested in the native application.
Syntax and documented data rules
=STEYX(known_y's, known_x's)
Supply dependent y-values first and independent x-values second. The official entry requires equal counts and explicitly documents #N/A for unequal numbers of data values.
It says blanks, text, logical values and errors contained in an array are ignored; it separately describes numeric text and logical values supplied directly as being interpreted numerically. Those statements should not be expanded into a guarantee that arbitrary malformed or one-sided missing pairs are safe. Use aligned, complete numeric observations for the reproducible models here.
The reference describes a standard error for the regression line. For ordinary simple linear regression with a fitted intercept and slope, the mathematical check is the square root of the residual sum of squares divided by nā2. The two fitted coefficients account for the two lost degrees of freedom. This is the residual-standard-deviation calculation in the NIST least-squares reference.
Rebuild every residual in a worksheet
Create a blank sheet named ResidualSpread. Put x and y in A1:B1, then enter:
| Row | A: x | B: y |
|---|---|---|
| 2 | 1 | 2 |
| 3 | 2 | 4 |
| 4 | 3 | 5 |
| 5 | 4 | 4 |
| 6 | 5 | 5 |
Use G for labels and enter the following in H:
- H2:
=STEYX(B2:B6,A2:A6) - H3:
=SLOPE(B2:B6,A2:A6) - H4:
=INTERCEPT(B2:B6,A2:A6)
Expected H3 and H4 are 0.6 and 2.2. Put Fitted y, Residual and Squared residual in C1:E1. Enter =$H$4+$H$3*A2 in C2, =B2-C2 in D2 and =D2^2 in E2. Fill C2:E2 through row 6.
| Row | Fitted C | Residual D | Squared residual E |
|---|---|---|---|
| 2 | 2.8 | -0.8 | 0.64 |
| 3 | 3.4 | 0.6 | 0.36 |
| 4 | 4 | 1 | 1 |
| 5 | 4.6 | -0.6 | 0.36 |
| 6 | 5.2 | -0.2 | 0.04 |
Enter =SUM(E2:E6) in H6; expect 2.4. Enter =SQRT(H6/(5-2)) in H7; expect 0.894427190999916, agreeing mathematically with H2. The literal 5 is the number of complete pairs in this fixture; update it if you deliberately change the dataset.
Why dividing by five gives a different measure
Put =SQRT(H6/5) in H8. Its expected result is 0.692820323. This is the square root of the average squared training residual, using denominator n. It differs from STEYX’s residual standard error for the fitted two-coefficient model.
Neither number is a percentage. If y measures hours, both are in hours; if y measures currency, both are in currency units. Do not confuse the residual spread with a correlation coefficient bounded between -1 and 1.
The one returned value also does not give the uncertainty for every possible prediction point. An interval for a fitted mean or a new observation needs additional assumptions and terms, including the location of the requested x. Simply adding and subtracting STEYX from a prediction is not a documented confidence-interval procedure.
Advanced: demonstrate the effect of measurement units
Keep the original model intact. Put Scaled y and Shifted y in J1:K1. Enter =B2*10 in J2 and =B2+100 in K2; fill J2:K2 through row 6.
Use N2 for =STEYX(J2:J6,A2:A6) and N3 for =STEYX(K2:K6,A2:A6). Independent targets are 8.944271910 and 0.894427191. Multiplying y by ten multiplies residuals and their standard error by ten. Adding a constant moves the fitted intercept while leaving residual spread unchanged.
For a separate perfect-line control, enter 1,2,3,4 in A2:A5 and 3,5,7,9 in B2:B5 on a new sheet named PerfectLine. Put =STEYX(B2:B5,A2:A5) in D2. The mathematical target is 0, because y=1+2x fits all four pairs exactly. This exact sample fit does not prove that future observations have no noise.
Reproduce the official screenshot
On a separate sheet named VendorSteyx, enter A1:B4 without headers as 1,0; 5,1; 2,12; 6,8, one pair per row. Here column A is known y and column B is known x.
Put =STEYX(A1:A4,B1:B4) in D2. ONLYOFFICE’s screenshot shows 2.904819. Independently, the residual sum of squares is approximately 16.875949367; dividing by 4ā2 and taking the square root gives 2.904819217.
Errors and related functions
Use more than two complete observations with variation in x for this fitted two-parameter model. With only two points the denominator nā2 is zero; constant x cannot determine a unique ordinary slope. Apart from its explicit unequal-count #N/A rule, the short entry does not enumerate every native error for these boundaries.
Keep residual calculations at stored precision, rather than recomputing them from displayed rounded predictions. SLOPE and INTERCEPT describe the line. RSQ describes squared correlation, and FORECAST.LINEAR evaluates the line at a chosen x. Each answers a different question.