Semi-supervised learning uses labelled and unlabelled examples together to learn a task. In its predictive form, some training examples have observed target labels, while others supply only inputs. For a quant classifying company announcements, it offers a way to use a document archive when expert annotation is scarce. More documents can inform the fit; they do not create more verified outcomes.
The defining distinction is that the unlabelled examples participate in learning. Fitting a supervised model on labelled documents and then merely predicting the rest is ordinary prediction. Semi-supervised methods let the unlabelled inputs influence the fitted predictor, for example through their similarity structure or model-generated provisional labels.
Ten documents, two observed labels
Consider a synthetic training pool of 2 labelled announcements and 8 unlabelled announcements, all available before a declared fitting cutoff. A human has classified one as earnings-related and the other as “other”. The remaining documents have usable text but no human task annotation.
An ordinary supervised baseline fits only the labelled examples. A semi-supervised alternative also uses the unlabelled text to constrain the fit. It might favour similar predictions for documents close in a chosen feature representation, or add selected model predictions as pseudo-labels during further fitting.
If all 8 unlabelled documents receive pseudo-labels, there are still only 2 human-labelled examples. A pseudo-label records what the model currently believes, not an independently established answer. The counts describe provenance; this tiny pool supplies no evidence of classification accuracy or trading value.
Synthetic company-announcement example. Each square is one document; filled squares have human labels and outlined squares have unknown task labels. Pseudo-labels can guide training without adding human annotations.
What can unlabelled data contribute?
Unlabelled inputs reveal where observations occur and how they resemble each other. They do not, by themselves, identify which regions mean “earnings-related”. The human labels anchor that meaning; an additional modelling assumption connects input structure to the task.
| Approach | How the unlabelled examples affect learning | Assumption to challenge |
|---|---|---|
| Self-training | Fit a classifier, assign selected confident predictions as pseudo-labels, and refit | Confidence must be informative about correctness; initial errors can reinforce themselves |
| Graph-based learning | Link similar examples and encourage compatible predictions along those links | Neighbourhoods in the representation should connect examples with similar task labels |
| Cluster-based learning | Use groups in the inputs to help distribute limited task labels | A cluster should be sufficiently consistent with the target classes |
These are alternative mechanisms, not steps every model must follow. In financial text, documents may cluster by issuer or writing style while the intended label concerns economic content. That structure can be irrelevant or misleading. Unlabelled data can therefore hurt as well as help. Zhu's author-written introduction explains why the relationship between input structure and the target matters; the scikit-learn guide describes self-training and graph-based propagation.
Unlabelled does not mean available for training
For a chronological financial experiment, both labelled and unlabelled training inputs must have been usable by the fit time. Human annotations used for fitting must also have been available then. Removing the labels from future documents does not make their text historical information. Fitting a representation, similarity graph or scaler on that future block can introduce look-ahead bias.
Choose pseudo-label thresholds and other settings inside the development process. Compare against the labelled-only baseline using genuine held-out labels, and reserve a later test period for the frozen procedure. Agreement with the model's own pseudo-labels is not independent validation. The Time-Ordered Train, Validation and Test Design lesson develops fitting cutoffs, completed-label eligibility and pipeline boundaries. Its public preview is available; the full lesson requires membership.
A deliberately transductive experiment is different: it uses the particular unlabelled collection whose labels are sought. That can be a valid task, but it does not establish the performance of a historical forecasting system on future arrivals. State which experiment is being evaluated.
Finally, self-supervised learning constructs a learning signal from the data itself, such as predicting hidden text. Semi-supervised learning describes the use of labelled and unlabelled data for a task. Self-supervised pretraining can be part of such a pipeline; the terms describe different aspects of learning.
Sources
- Xiaojin Zhu, Semi-Supervised Learning, author-hosted encyclopedia entry, pp. 1–4. Definition, inductive versus transductive learning, and the assumptions connecting unlabelled structure to prediction.
- scikit-learn: Semi-supervised learning, checked 27 October 2026. Self-training, confidence and graph-based label propagation.
- scikit-learn: Common pitfalls—data leakage, checked 27 October 2026. Training boundaries for fitted preprocessing.
- Aurélien Géron, Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, second edition (O'Reilly, 2019), Chapter 9, “Using Clustering for Semi-Supervised Learning”. Read in the local text; the publisher's edition page identifies the edition. Cluster-based propagation and its sensitivity to incorrect propagated labels. The document-count example here is original and synthetic.