Sharing information on market, production capacity and inventory statuses is well known tohelp supply chain members reduce the inventory costs and mitigate the bullwhip effects.However, in practice, many manufacturers still have difficulties in obtaining market andinventory information from their customers due to the lack of trust, not major account, highinvestment cost, etc. When information sharing is not feasible, manufacturers can resort toforecasting to predict the plausible next order arrival time and quantity so that they can pre-produce to shorten order lead time without carrying excessive inventory.With hybrid MTS and MTO production systems, manufacturers can benefit by implementingpostponement to pre-purchase production parts and / or pre-produce MTS strategy based onthe forecasted order arrival time and quantity. In this study, we consider forecasting the nextorder arrival time and order quantity under the hybrid MTS and MTO production environmentwhen the manufacturer has no information sharing with his customers. Most forecastingsystems rely on complicated rules to enhance their accuracy; however, the rules are notsimple enough for decision makers to convert them into human intelligence to make areasonable good forecast. We adopt the adaptive neural-fuzzy inference systems (ANFIS) asthe fundamental inference system to infer the next order arrival time and order quantity aftertraining with previous order related information to obtain a fuzzy-inference-system (FIS). AsANFIS can predict the next order arrival time and quantity, manufacturers or suppliers areable to pre-purchase or pre-produce customers’ orders to reduce order lead times. Toprovide decision makers with tractable decision domains and inference rules, we furtherrestrict the levels of fuzzy decision variables to have small amount of fuzzy inference rules.By learning the simple rules provided by the ANFIS, decision makers can still make properpurchase / production decisions by observing the statuses of the supply chain without havingto perform the ANFIS.
We experimented with sinusoidal market demand patterns and found that an ANFIS with 10x 10 rules can have MAPE (mean absolute percentage error) of 1% for order arrival time andquantity as well. By reducing the ANFIS to 3 x 5 rules, the ANFIS can still have MAPE of approximate 15%, which is accurate enough for decision makers to learn the 3 x 5 rules andmake decisions without relying on the ANFIS to achieve a reasonable good forecasting result.