Airline Empires - Airline and Aviation Articles, Airline Simulation - New Methods to Forecast Passenger Demand. Part 1: Page 3

New Methods to Forecast Passenger Demand. Part 1

Page 3 of 4

A neural network is a method of recognizing patterns in a set of data, and using those patterns to forecast future results. It mimics the human brain, a very able pattern recognizer, and uses different nodes with which to calculate and test data. These nodes are referred to as neurons, which complete the human brain analogy. The neural network requires input variables, which are linked to a hidden layer of nodes where manipulation of the variables takes place. The output variable (or variables) is the dependent variable, and is the result for the set. In order to accurately forecast with a neural network, it must first be trained. Training a neural network is simply a means of supplying the network with both the input variables and the output variable for several data sets, and manipulating the data from the input variables until the relationship between each is sufficiently found. Once the network has been properly trained, it can accurately predict a dependent variable, considering accurate independent variables are entered into the input layer. For instance, to predict how the DOW industrial average would react to a certain mix of events, it must first be determined what variables affect the stock market. It is impossible to identify every variable, so it is important that as many variables as can be found be used in order to provide accurate data. Then, after entering the historical data of each variable into the neural network as the input nodes, and the resulting DOW value, researchers can proceed with training the neural network. Once this has been completed, other values can be added into the independent variables and predict how the output variable will react. Even though the human brain is extremely efficient at the pattern recognition required for neural networks, it cannot handle the massive amounts of data required for accuracy. Because of this, neural networks are always created by computer software, and the ability to train the networks lends the software the “artificial intelligence” title.

Direct application of neural networks to airline industry forecasting was made when BaFail (2004) introduced neural network forecasting to domestic and international passengers at specific markets in Saudi Arabia. BaFail (2004) used 15 factors that he considered vital to the prediction of passenger travel, and was able to achieve mean absolute percentage errors between 1.2% and 31% for domestic travelers, and between .3% and 22% for international travelers (BaFail, 2004). The bulk of the percentage errors were below 10%, and this study illustrated the need for a comprehensive set of factors to be included in order for a more accurate forecast to be achieved.

Using neural networks to forecast information as inconsistent as market profitability in non-stop airline markets has both advantages and disadvantages. According to Law (1998), who used neural networks in the previously mentioned study to predict hotel room occupancy rates, a neural network is superior to other methods when the data set is small, has large amounts of missing data, or is noisy (inconsistent). Van Wezel and Baets (1995) found that some researchers consider the neural network as more of a “black box” approach, where data goes in and magically comes out without much understanding of how it is derived. While this argument does seem valid, it can be argued that much has happened over the last 10 years in regards to researchers’ perception and trust in computer programs, and this “black box” is now more easily understood. It is understandable that a researcher in 1995 who had been working with simpler regression techniques would be skeptical of software that claims to exceed proven forecasting methods without much explanation other than “pattern recognition” in how it achieves such strong results. Van Wezel and Baets (1995) also are quick to point out the susceptibility of neural networks to the quality of data being used. This limitation is two-fold, since not only the data must be accurate, but in order to achieve accurate results, researchers need to include a comprehensive set of independent variables that will affect the dependent variable in any way. Heuglin and Vannotti (2001) applied neural network techniques to sorting through extremely large banks of data with promising results. Further research into marketing applications of neural networks shows that in data mining applications, using a back-propagation neural network model and assigning weights to each variable yields even more accurate results (Fish & Segall, 2004).

Even though there is seemingly unlimited research regarding the factors that affect market profitability, using neural networks to forecast it has yet to be thoroughly explored.

HYPOTHESIS
The use of neural networks in predicting market response to several different factors will be tested to produce mean absolute errors consistently below 10%, which will allow for accurate forecasting in several different types of non-stop markets.

<< Start < Previous 1 2 3 4 Next > End >>