A Robust Forecasting Framework for Multi-Series Cash Flow Prediction

Cash flow forecasting is crucial for decision making in micro-economics and macro-economics activities. Typically, cash flow series show high volatility, high diversity, and complex seasonality. Moreover, series are generally not sampled frequently enough to have sufficient training data for the powerful deep models. To overcome the above challenges, we propose a robust forecasting framework which firstly regulates series with outlier smoothing and a linear seasonal adjustment model, then uses an ensemble technique to exploit advantages of both statistical and deep time series models. To solve the data scarcity problem in training deep models, we propose a cross learning approach which train a single model using data from multiple series. We validate the proposed framework in a real dataset. Experiments demonstrated that the proposed forecasting framework leaded to superior performance than state-of-the-art models.