A1971
Title: A semi-parametric additive hybrid model: An application to SSE composite index forecasting
Authors: Bolun Qu - Monash University Malaysia (Malaysia) [presenting]
Erniel Barrios - Monash University Malaysia (Malaysia)
HowChinh Lee - Monash University Malaysia (Malaysia)
Siew Eu-Gene - Monash University Malaysia (Malaysia)
Abstract: Financial time series forecasting confronts a trade-off between the interpretability of regime-switching econometric models and the predictive capacity of deep learning. Existing hybrid architectures typically rely on globally linear baselines and static sequential estimation. This propagates specification errors into the neural network and neglects the skewed empirical distribution of market regimes. A regime-switching semiparametric additive hybrid model is proposed that integrates a threshold autoregressive process with a long short-term memory network, jointly estimated by a decoupled iterative backfitting algorithm. The model replaces the globally linear baseline with a regime-dependent parametric process, allowing the threshold mechanism to extract state-conditioned linear memory before the residual is mapped by the recurrent network. An inverse-frequency data cloning mechanism mitigates regime imbalance while preserving the temporal topology required by the recurrent architecture. The balanced cloned sample feeds only the pointwise TAR component, while the LSTM receives chronologically ordered observations. The model is evaluated on the Shanghai Stock Exchange Composite Index for joint forecasting of closing prices and logarithmic returns. Out-of-sample results indicate improved point forecasting accuracy relative to linear benchmarks, alongside statistically significant directional predictability by the Pesaran-Timmermann test.
