Predicting 1H shifts before you measure: what the confidence interval tells you
Author: Dr. Oliver Kraft, Dr. rer. nat., Organische Chemie · Updated 2026-08-04
Short and direct
A shift prediction does not replace a measurement. It answers the question before it: which signals will overlap, and which will carry an assignment. For paracetamol the tool predicts nine proton signals. The NH signal sits at 9.24 ppm with an interval only one ppm wide, whereas the aromatic H-C4 sits at 7.29 ppm with an interval from 4.54 to 10.04 ppm.
That is the actual information. An assignment can rest on the NH signal, not on H-C4. Reading the prediction as a point value and ignoring the interval width treats both alike and justifies an assignment with a number the model itself does not support.
The computed result
| Position | Predicted (ppm) | Confidence interval | Width (ppm) |
|---|---|---|---|
| H-N1 (Amid-NH) | 9.24 | 8,74 bis 9,74 | 1 |
| H-O2 (phenolisches OH) | 8.84 | 7,10 bis 10,59 | 3.49 |
| H-C4 (aromatisch) | 7.29 | 4,54 bis 10,04 | 5.5 |
| H-C5 (aromatisch) | 7.28 | 6,66 bis 7,90 | 1.24 |
| H-C2 (aromatisch) | 6.65 | 6,44 bis 6,86 | 0.42 |
| H-C3 (aromatisch) | 6.63 | 6,21 bis 7,04 | 0.83 |
| H-C1a (Acetyl-CH3) | 1.95 | -0,80 bis 4,69 | 5.49 |
| H-C1b (Acetyl-CH3) | 1.95 | -0,80 bis 4,69 | 5.49 |
| H-C1c (Acetyl-CH3) | 1.94 | -0,81 bis 4,68 | 5.49 |
Computed 2026-08-03 with covnmr_predict, nucleus 1H, SMILES CC(=O)Nc1ccc(O)cc1. Verbatim tool output, method csp5.
What is a prediction actually good for?
For three things. First, planning: if two expected signals fall in the same region, a different solvent or a two-dimensional experiment is worth it, and that is better decided before instrument time than after. Second, cross-checking: a measured signal that deviates clearly from the predicted range points to a different structure, a by-product, or a solvent effect. Third, candidate testing when several structure proposals fit one mass.
What it is not good for: as evidence. A predicted shift is a model expectation, not a measurement. In a submission the measured spectrum appears, not the prediction.
Why are the three methyl protons not identical?
In the result below the acetyl methyl group carries three values, 1.95, 1.95 and 1.94 ppm. Chemically these three protons are equivalent through free rotation and give a single signal. The three slightly different numbers are an artefact of the per-atom prediction, not a physical finding. A control run confirms this: for toluene the same method predicts exactly 2.36 ppm for all three methyl protons. The model does not average over rotationally equivalent nuclei, it predicts each nucleus individually, and whether the values coincide is a numerical outcome, not intent. Copying the output into a table unexamined documents a difference that does not exist.
A second point belongs here, and it is the more common source of error: the labels H-C2 to H-C5 are the tool's internal atom indices, not IUPAC position numbers. A control run on phenol shows this clearly. There H-C2 sits at 7.22 ppm, in the range of the meta protons, while the ortho protons appear as H-C4 and H-C5 at 6.81 and 6.82 ppm. Reading H-C2 as the ortho position to the hydroxyl group, which IUPAC numbering would suggest, assigns it wrongly. This page therefore quotes the labels exactly as the tool emits them and derives no chemical position from them.
That is why this page shows the raw output rather than a tidied version. A prediction is a tool for someone who knows what they are reading.
Frequent questions
- Does the prediction account for the solvent?
- The run shown passes no solvent. Exchangeable protons in particular, NH and OH, shift considerably with solvent and water content and can disappear entirely. The wide intervals on those two signals reflect exactly that uncertainty.
- Why does the methyl interval reach below zero?
- Because the model applies a very wide expectation range for aliphatic protons and the lower bound falls below zero purely arithmetically. Physically, negative shifts relative to tetramethylsilane are possible, but not expected for an acetyl group. That is a model limitation worth knowing.
- Can I predict 13C as well?
- Yes, the nucleus is a call parameter. For structure elucidation the combination of both nuclei is considerably more informative than either alone, because the ambiguities of the two ranges rarely coincide.
- How do I compare prediction and measurement properly?
- Not signal by signal by smallest distance, but by assignment. Two signals whose intervals overlap must not be assigned separately on the strength of the prediction alone. Two-dimensional experiments exist for that, as does a tool for comparing predicted and measured spectra.
- What does the method label csp5 mean?
- It names the prediction method used and appears in every row of the output. It belongs in the documentation whenever the prediction appears in a report: stating which method produced a number is part of traceability.
Sources
- ICH Q6A: Specifications and test procedures for new drug substances and products. Frames the question of which analytical statement must be evidenced.
The same calculation on your structures
The results above come from the same tools your language model can call over the Model Context Protocol. Test them free of charge, at your own pace.