Enhancing Innovation Diffusion Models with Neural Residual Modeling for Improved Forecasting

Innovation diffusion models such as the Bass model are widely used to describe long-term adoption patterns but often fail to capture short-term fluctuations and volatility. This thesis proposes a hybrid forecasting approach that enhances diffusion models through residual modeling, combining structural diffusion with statistical and neural correction models. Classical and dynamic market potential Bass models are paired with SARIMAX and neural residual models, including Temporal Convolutional Networks and multi-horizon TCNs. The approach is evaluated using battery electric vehicle adoption data for Italy and Norway. Results show that residual modeling consistently improves forecast accuracy, with neural and multi-horizon approaches providing the most stable long-term forecasts. Dynamic market potential improves performance in mature and stable markets, while simpler diffusion baselines remain more robust in volatile settings. Overall, the findings demonstrate that neural residual modeling effectively enhances innovation diffusion forecasts when matched to market conditions and forecasting objectives.