In almost all plans, there is forecasting. In daily life and business, we need to make decisions that involve uncertainty. The most common causes are probably the weather and financial forecasts. To be able to forecast the exchange rate or the stock market index will be financially rewarding. Apart from these, you might need to predict sales volume for inventory management, expenses to prepare a business budget, or the traffic to decide when to leave home to arrive at the airport on time. Forecasting can be defined as “an attempt to determine the likely value of a variable."1 The critical word in this definition is "likely." There is a random factor in most of the processes that generate the variable we are interested in. This random component of the variable is called noise, whereas the element that comprises the processes that form the features of the variable is called signal. Therefore, even if you make a point forecast, what we have is a likely value of the variable calculated using “estimated probability distribution of the variable to be observed in the future”.2
While good forecasts can be rewarding, wrong forecasts may lead to poor decisions and consequential losses. One of the most frequently quoted poor predictions was made by Yale economist Irving Fisher. He argued that “Stock prices have reached what looks like a permanently high plateau” a few days before the stock market collapse in October 1929.
Since forecasting is essential, especially for business management, what is the recipe for a good forecast? First, some variables are almost impossible to forecast. For instance, you can't predict the outcome of a coin toss. If it is a fair coin, the probabilities of heads and tails are equal, and the result is entirely random. A common mistake made by gamblers is ignoring that the outcome of a coin toss is independent of previous results of the coin tosses. Having ten consecutive heads doesn't mean that the probability of having a head again is lower than fifty percent. Therefore, historical data is useless, and there is no signal that we can use to forecast. On the other hand, we can predict the time accurately that the sun will rise tomorrow since there is no noise in this case.
Most of the variables we are interested in will be between these two extreme points of zero noise and zero signal. You can only forecast the change in the signal component of the variable.
Secondly, you need to know the features of the variable that make up the signal to construct the forecasting model. It is easier to figure out these features for some problems, especially if there are theories or previous studies about the issue you are interested in. For instance, there is a ton of literature on the prediction of bankruptcy to determine the features of credibility if you are interested in developing a credit score for a bank's customers. Third, you need lots of data to estimate the target variable's sensitivity to the features you select.
Finally, the forecasts and the actions taken by other parties today can change future outcomes. The weather forecast is free of such problems since no human intervention can change the weather tomorrow. However, financial markets are open to government interventions. Too big to fail is a prevalent example of government intervention. You can forecast a bankruptcy or a market collapse, which can be prevented by the support of other parties to relieve the adverse effects of the forecasted event.
Forecasting is a challenging task, especially for variables that involve hidden and unobservable features. For most cases, you will not have a point forecast with no uncertainty. You will have an estimated distribution that represents several likely outcomes and their probabilities.
Prof. Dr. Cenktan Özyıldırım
References:
1. Brooks, Chris. (3008) Introductory Econometrics for Finance., Cambridge: Cambridge University Press
2. Hyndman, R.J., & Athanasopoulos, G. (2018) Forecasting: principles and practice, OTexts: Melbourne, Australia