My undergraduate thesis explored whether nonlinear models could capture relationships between company characteristics and asset prices in the Chinese A-share market — and how a model could account for both cross-sectional information and time.
The research question
Traditional asset-pricing approaches often make linear assumptions. This project examined how machine-learning methods could represent nonlinear relationships, using data on CSI 300 equities and 68 company characteristics.
The research considered the period from January 2010 to January 2022. It compared conventional approaches with machine-learning models, and examined the relative importance of the input features.
Methods & implementation
I implemented and compared support vector machines, gradient-boosted regression trees, random forests, and multilayer perceptrons alongside linear and recurrent-model benchmarks.
The project also explored a hybrid architecture combining convolutional feature extraction, LSTM sequence modelling, and an attention mechanism. The motivation was to bring feature interaction and temporal context into the same model.
- Collect and preprocess company characteristics and equity data.
- Separate training, validation, and test samples for model development and evaluation.
- Compare out-of-sample fit across model families.
- Examine feature importance and its stability over time.
What the study found
Within the historical experiment, nonlinear models generally fitted the data better than the linear benchmark. Increasing neural-network depth did not consistently help; more complex models could overfit.
The hybrid model performed best in the thesis’s comparison. Feature-importance analysis also suggested that a relatively small subset of characteristics explained much of the model’s behaviour. These are findings from that academic experiment, rather than claims about live trading performance.
Limitations worth keeping
The thesis explicitly notes a potential look-ahead issue: the financial database may contain adjusted company data rather than a fully point-in-time history. This limits the conclusions that can be drawn from the historical results.
Other constraints include model interpretability, limited computing resources, and randomness across training runs. A stronger follow-up would prioritise point-in-time data, robust out-of-sample evaluation, and a clearer economic explanation of the model’s behaviour.
Why it still matters to me
This was an early meeting point between my engineering background and my interest in investing. It brought the modelling problem and the data-quality problem into the same frame — a useful perspective for the research tools I build today.