Market Capitalization Forecasting

The task

Forecasting firm-level market capitalization at one, two and three year horizons from 25 years of historical market data. Longer horizons are substantially harder — the signal degrades quickly, so a model that looks strong at one year can be near-useless at three.

Why three architectures

An MLP treats the input as a flat feature vector and ignores sequence order entirely, which makes it a useful baseline: if a recurrent model cannot beat it, the temporal structure is not being exploited.

An LSTM models the sequence directly, carrying state across time steps. An encoder–decoder separates reading the history from generating the forecast, which suits multi-step horizons better than forcing a single model to do both at once.

Comparing all three answers whether the added complexity actually earns its place at each horizon.

Dimensionality reduction

Twenty-eight financial indicators were compressed with PCA before modelling. With a limited number of firm-years available and highly correlated indicators, the raw feature set invites overfitting — reducing to the principal components that carry most of the variance improved generalization.