Ametek.txt - Price forecast of a listed stock
Red line=Target output, Average Price Per Share (Avg$/Sh)
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of Average Price Per Share (Avg$/Sh)
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: The genetic algorithm (GA) based forecaster net Ametek.ga was able to detect the change in direction beyond the end of the training range (i.e., the 3 blue empty boxes after the learning range). The unseen data (i.e., the two red empty boxes after the learning range) confirmed that the directional change was correct.

Chaos.txt - A classical example on chaotic time-series forecast
Red line=Target output, chaotic time series
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of chaotic time series
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This is a classical chaotic time series prediction problem based on the Mackey-Glass equation. The neural network was able to figure out the underlying equation quickly and provided future values of the time series.

Citicorp.txt - Citicorp stock price forecast
Red line=Target output, closing price
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of closing price
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This is a stock price prediction based on high, low, close prices and volume of the stock. Prices and volume alone usually do not provide good predictive information, but in this case, the GA based network was able to do a modest job. The upward and downward arrows indicate the optimistic and pessimistic range of the forecast.

Credit.txt - Credit/Bank loan applicant classification and screening
Red line=Target output, rating of 100 means approved, 0 means rejected
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Classifiction of applicants
Blue solid box=Classifiction in learning range (old cases)
Blue empty box=Classifiction of new cases

Remark: This is a typcial classification problem. Each of the records (representing the applicants) was assigned a value by the neural network, indicating the outcome of the classification. The classification value can be used as it is (e.g., value > 80 means good rating, application approved, > 50 means marginal and requires further consideration, and < 50 means rejected), or with a hard limit function (e.g., >50 means approved, while <= 50 means rejected). Classification problems are much easier to solve than time-series problems, and many other similar problems can be built using this example.

Djia.txt - Dow Jones Industry Average forecast
Red line=Target output, DJIA
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of DJIA
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This application demonstrates the combined power of neural network and genetic algorithms. A population of neural networks were first created, each was assigned to forecast the time series with different time horizons (e.g. some forecast 1 step ahead, some 3 steps ahead, some 7 steps ahead, some 15 or even 50 steps ahead). After generations of evolution and competition, the best network will evolve. In this example, the best performed a 3-step ahead forecast.

Users can choose the range of the horizon (e.g. from 1 to 50 steps ahead) and let GA determine the best horizon to use. Or they can fix the range (e.g. at 5 steps ahead) which usually does not provide optimal results. A better approach is to first choose a desired range (e.g. from 1 to 90 steps ahead, or even longer if you have a lot of data), let GA determine the best horizon (say, if 50 steps ahead was chosen), and then fix the range at the best horizon found (i.e., 50 steps ahead in this case) and retrain the network to forecast at this horizon.

Estate.txt - Forecast of quarterly property demand
Red line=Target output, quarterly housing property demand
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of quarterly housing property demand
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This example illustrates another useful feature of the program, to avoid over-training and hence over-fitting. Both the learning and testing ranges can be partitioned such that only certain amounts of the data will be used in the learning and testing process. In the above example, only 3/4 of the learning range and 1/4 of the testing range were used. The forecast showed a 3-quarter projections beyond the data range.

Property.txt - Valuation of property prices based on various factors
Red line=Target output, property prices
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Estimated property prices
Blue solid box= Estimated property prices in learning range
Blue empty box=Estimated property prices

Remark: This is another classification problem in which each data point represents a property to be valuated, and the neural network generated a classification value for each property indicating the estimated value of it. Input data includes location, area, facilities and etc.

Sinewave.txt - A classical example on sine wave prediction
Red line=Target output, sinewave pattern
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of next sinewave pattern
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This is a classical sine wave problem for neural network computing. A number of sine waves with different frequencies were mixed and used to train the neural network.

Sp500.txt - S&P500 forecast
Red line=Target output, S&P500 index
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of S&P500
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: The network trained to forecast the S&P500 did a marvellous job in detecting the upward turn (the 5 blue empty boxes). Download an evaluation copy now and see it for yourself!

Sports.txt - Soccer game prediction
Red line=Target output, relative ranking of soccer teams
Red solid box=Learning range (1.0 means best team, 0 means worst team)
Red empty box=Unseen data (ranking equals 0.5)
Blue line=Predicted relative ranking
Blue solid box=Prediction in learning range (past matches)
Blue empty box=Prediction for new match

Remark: This is an unusual application which requires some ingenuity in setting up the training file. The network was trained with past performances of the teams, and was able to pick the winning team (the team with a score of 1.0) for the upcoming match, and ranked other teams accordingly (1.0 = best, 0.0 = worst).

Stock6m.txt - Stock market 6-monthly return forecast
Red line=Target output, 6-monthly returns
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of 6-monthly returns
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: In this example, each data point represents a 6-monthly return from the stock market. The capability of the neural network in detecting the market turning points were again clearly illustrated.

Stocksel.txt - Stock selection/ranking
Red line=Target output, annual returns for ranking/selection
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of stock return
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: The neural network was trained to forecast the annual returns of various stocks based on several company ratios, and displayed the results in bar chart format. Each bar represents a stock, the performance of which can be easily seen from the chart display. Similar applications include bond rating, company bankruptcy predictions and etc.

Us$_dm.txt - USD/DM exchange rate forecast
Red line=Target output, US to DM exchange rate
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of USD/DM exchange rate
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This application example demonstrates a useful feature which excludes irrelevant data regions from the training process. Those historical data points on the left of the chart were in the VP (Very Positive) and P (Positive) regions in the data distribution, and were therefore excluded (indicated by the darkened area). The more recent data points were mainly in the M (Moderate), N (Negative) and VN (Very Negative) regions, so only those in these regions were used. This method of excluding irrelevant data speeds up the training and provides a better accuracy.

Us$_dm2.txt - USD/DM exchange rate forecast
Red line=Target output, US to DM exchange rate
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of USD/DM exchange rate
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: This is another USD/DM exchange rate forecast. Users can use the Zoom-In buttons provided on the top-left corner of the chart to see the details and level of accuracy of the forecast.

Us$_yen.txt - USD/YEN exchange rate forecast
Red line=Target output, USD/YEN daily exchange rate
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of USD/YEN daily exchange rate
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: Users can also test the predictive power of neural networks on short-term (e.g. daily) exchange rate forecast. In this USD/YEN application, various cross rates were used as inputs, and as demonstrated, the downward and upward turns were detected.

Us$_sgd.txt - USD/SGD 6-month ahead forecast
Red line=Target output, USD/SGD monthly exchange rate
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Forecast of USD/SGD 6-month ahead
Blue solid box=Forecast in learning range
Blue empty box=Forecast

Remark: In this USD/SGD exchange rate forecast, each blue box represents a 6-month ahead forecast. The network was able to pick up the various crucial turning points, and generated 6 monthly forecasts beyond the last data point. The downward turn was later proved to be correct.

Xor.txt - A classical example on XOR
Red line=Target output, 1 if both inputs differ, 0 if equal
Red solid box=Learning range
Red empty box=Unseen data
Blue line=Exclusive OR
Blue solid box=Exclusive OR in learning range
Blue empty box=Exclusive OR

Remark: This is another classical example to test the basic capability of neural networks. Compared to time-seris and classification problems, this one was found to be extremely fast and easy. Many network tools are not even able to solve this problem with the same level of efficiency and speed as NeuroForecaster, let alone other more sophisticated tasks.