Recommended links

PSM++ Simulation System

Simulation Encyclopedia

Queuing models


A great small prediction tool


It has always been a great challenge for men to be able to predict what will happen in the future. The changes of weather, earthquakes, diseases and harvests have been crucial for our behavior and actions. So, the men have always been looking for some clues and ways to predict future events and to take appropriate actions. Recently, this mainly means money. The price of oil and other goods and the patterns in the stock exchange are subject to frenetic search for prediction methods, which, if they exist at all, may make someone rich without much effort. 

There are two sources of information that can help us to predict something: The historical data and our knowledge about the structure and dynamics of the system we observe. These two sources should be combined and used in a comprehensive way to avoid heavy errors. If we base our prediction on the historical data only, we can conclude, for example, that beating tam-tams is the best way to make the solar eclipse terminate. Other example of misinterpretation of the historical data is the following. Suppose I want to estimate the probability that I die during the next year period, using the observations of all my life. The result is quite optimistic: the probability of such event is equal to zero, because my observations tell me that always other persons dye, never me.

The analysis of historical data has been subject to intensive research and some more reasonable methods have been developed. These methods look for the patterns in the past behavior of a time-series (historical records). The tendencies are extracted from the data. These patterns and tendencies are used to calculate possible future system behavior. This, in some cases may provide useful prediction methods.

The other approach to the predictor problem is to understand the system structure and its dynamics and to reproduce its behavior for a future time interval. This is what we call modeling and simulation. Using a computer as the main simulation tool can provide powerful prediction methods. Observe, however, that a real physical, social, economic or financial system cannot be modeled in an ideal way. Our knowledge on its structure and functioning is always limited, and even a small logical error (what “small” means?) in the modeling process can result in completely erroneous simulations.

Undoubtedly, the two approaches mentioned above (time-series and simulation) should be combined to achieve our goal. However, if we cannot apply one of them, then we should use the other, with appropriate precaution.

The PredictHit program is a time-series analyzer. The input to the program is a sequence of values, supposed to be the values of a variable of interest taken from a stochastic process. The values must be given for equal time intervals. These can be, for example the demand for a product for consecutive days, weeks, months etc. during some past time interval. PredictHit looks for tendencies in the series and for possible seasonal (periodical) changes. Once an approximating curve is found, a prediction for the future is being produced. If the provided data is charged with random errors or uncertainty, you can get the predictions with corresponding confidence intervals for the prediction value, as function of time.

The data can be typed and stored in a file, retrieved, edited or imported from any other software that can export the data as an ASCII file.

PredictHit runs on a PC with Windows 98 or later, NT and XP, 64Kb of RAM, preferable 1024x768 screen resolution. At least 200 MHz machine recommended.

Here is an example of a data set to analyze (the image may be of poor quality because of the size reduction with respect to the original whole secreen plot). The horizontal axis is the time.

The following image is a fragment of a PredictHit screen. The red line is a final part of the analyzed data set. The prediction curve is shown in blue. These data has a periodic (seasonal) component that can be seen on both curves. This is a reduced fragment of the screen only, to save the display space.

As mentioned above, the data to be analyzed almost always is charged with some uncertainty. PredictHit offers the uncertainty analysis and shows how this afects the fitting curve and the prediction.

To carry out this analysis, you are asked to give the standard deviation for your data, and the sample size S. You also must give the confidence level, between 0.5 and 1. Then, S predictions are generated, each one on different data. In each prediction, each data point is changed randomly due to the given standard deviation. After generating S curves, the corresponding standard deviation for each time instant is calculated. The plot is shown with three curves: the original prediction P(time) (with unchanged data), and two curves. The probability of the prediction value to be inside the red region (between the two curves) is equal to the confidence level you gave before.

The following figure shows an example of a PredictHit screen with results of the uncertainty analysis.

A part of the past time instants and the prediction period are shown. The red region is the confidence interval for the fitting curve, when the data is subject to stochastic changes with given standard deviation. The horizontal axis is the time. With the probability defined by the user the predicted value is supposed to be inside the red region.

Contact: Stanislaw Raczynski 

Use this button to buy (US $24)

Get a GoStats hit counter