Anzeige
Erschienen in:
2013 | OriginalPaper | Buchkapitel
9. Planning Purchase Decisions with Advanced Neural Networks
verfasst von : Hans Georg Zimmermann, Ralph Grothmann, Hans-Jörg von Mettenheim
Erschienen in: Business Intelligence and Performance Management
Verlag: Springer London
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Abstract
In this chapter we investigate a typical situation of a corporate treasurer: on an ongoing basis some kind of transaction is performed. This may be a regular monthly investment in equities for a pension plan, or a fixed income placement. It might be a foreign exchange transaction to pay monthly costs in another currency. Or it could be the monthly supply of some commodity, like fuel or metal.
All these cases have in common that the treasurer has to choose an appropriate time for the transaction. This is the day on which the price is the most favorable. Ideally, we want to buy at the lowest price within the month, and we also want to invest our money at the highest available interest rate.
This problem is complex, because the underlying financial time series are not moving independently. Rather, they are interconnected. In order to truly understand our time series of choice, we have to model other influences as well: equities, currencies, interest rates, commodities, and so on. To achieve this we present a novel recurrent neural network approach: Historically Consistent Neural Networks (HCNN). HCNNs allow to model dynamics of entire markets using a state space equation: s
t+1=tanh(W⋅s
t
). Here, W represents a weight matrix and s
t
the state of our dynamic system at time t. This iterative formulation easily produces multi step forecasts for several time points into the future.
We analyze monthly purchasing decisions for a market of 25 financial time series. This market approximates a world market: it includes various asset classes from Europe, the US, and Asia. Our benchmar, an averaging strategy, shows that using HCNNs to forecast an entry point for ongoing investments results in better prices for every time series in the sample.