Performance-Aware Deep Learning for Stock Trend Prediction and Investment Decision-Making

Abstract

Deep learning methods are increasingly being applied to stock market prediction, with growing evidence that they can extract useful structure from financial time-series and related textual signals. However, prediction alone is not sufficient. In practice, model outputs must also be translated into sensible capital allocation decisions if they are to generate meaningful returns on investments. The framework integrates historical price data with text-derived features and evaluates multiple deep learning architectures across binary, regression and ordinal multi-class forecasting tasks. It further incorporates prior model predictions and their realised errors as meta-features, allowing the system to learn when its own outputs are more or less reliable. In parallel, a variable-horizon design is introduces to compare short-term and longer-horizon forecasting behaviour. Beyond prediction, the dissertation develops and evaluates a downstream staking and allocation layer using Kelly Criterion, Modern Portfolio Theory (MPT) and a hybrid strategy. Results show that this approach improves decision stability and supports stronger economic performance than accuracy only evaluation would suggest. In particular, longer-horizon classifi- cation models produced the most useful trading signals, while reliability-aware meta-features improved performance and risk adjusted return. The strongest investment result was achieved by the binary meta-feature system using MPT allocation, which returned 12.47% over 100 trading days. Overall, the finding show that financially useful stock prediction systems should be assessed not only by forecast accuracy but by how well calibrated and reliability-aware predictions translate into profitable and risk-conscious capital allocation.