Pooled Training Enables Stock-Agnostic Forecasting in the Universal Model

Summary

The document describes the central idea behind the Universal Model associated with Justin Sirignano and Rama Cont: train a forecasting model on pooled data from many stocks, then apply the shared model to forecast individual stocks, including stocks outside the training set. This differs from fitting a separate time-series model for each security. The model's claimed universality refers to this cross-stock applicability, rather than to a guarantee that every market or forecasting problem can use one unchanged model.

The response says pooled training performed better than stock-by-stock training in the discussed work and notes that other research replicated the finding. It does not explain the model architecture, define its “universal features,” provide detailed evidence, or establish that a small sample of stocks would reproduce the reported behavior. The exchange therefore conveys the pooling concept but leaves data requirements and generalization limits unresolved.

Key ideas

• The Universal Model is trained on pooled observations across many stocks rather than on one stock at a time.
• A shared model trained on pooled data is presented as applicable to stocks beyond its training set.
• The cited account reports better forecasting from pooled training than from separate stock-specific training.
• The exchange does not specify how much data is needed or whether a small stock sample would produce comparable results.
• The response describes the main modeling idea but does not define universal features or explain the model architecture.

Tags

# The "Universal Model" by Justin Sirignano and Rama Cont


# The "Universal Model" by Justin Sirignano and Rama Cont












In the nicely written article https://arxiv.org/abs/1803.06917 by Justin Sirignano and Rama Cont, they explained that their model is universal and stationary. I am a bit confused about some questions.

- What makes that model any different from other models?

- What do they called the "Universal Features"?

- As they use petabyte of data, i.e. 1000 stocks over on 3 years, can we achieve the same results with 3 stocks over 3 years instead?

## Answer by XiaoWang (score 2)

https://quant.stackexchange.com/a/50620

Its called 'universal' because, unlike usual models trained on time series for a given stock/ contract, this model is trained on a POOLED data set (in this case 500 or so stocks) and is then shown to be applicable for forecasting any stock, including those not included in the training data. This is different from the usual approach where, say, you use time series of IBM prices to estimate/train a model for IBM, then data for GOOGL to train/estimate a modle for GOOGL etc. Here they pool all data, train then use the model to forecast any stock. At first glance it seems nonsense but amazingly it works better than without pooling.

I think thats the main point of the paper.

This 'universality' and the superiority of training on pooled data has been confirmed by other papers which have replicated their results (for ex. Zohren et al)

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.