Analysis · Finance
How a Forecast Can Claim 100% Accuracy and Still Have Errors
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
A forecasting model scored 100% accuracy while still producing measurable prediction errors. That combination appears in Sandy Putra Siregar and Anjar Wanto’s 2017 study of North Sumatra’s Human Development Index. The revealing detail is what the paper leaves unclear: exactly how an error became an accuracy score.
The researchers used official data for 33 regencies and cities from 2011–2015. They trained neural networks using three annual observations to predict the following year: 2011–2013 inputs targeted 2014, while testing shifted the inputs to 2012–2014 and the target to 2015. Backpropagation repeatedly adjusted network weights to reduce prediction errors.
Five network configurations differed in the number of hidden neurons. Both the 16-neuron and 48-neuron versions received 100% accuracy scores. The authors preferred the latter because it required fewer training iterations, took less time and had a lower testing mean squared error. That error was approximately 0.000639 on normalized data, so the forecasts were not exact matches.
The paper lists an error range of 0.001–0.05 but does not clearly explain how this produces the reported accuracy percentages. Without that rule, 100% cannot be interpreted as perfect forecasting. Nor does the comparison establish an advantage over simply carrying forward the previous year’s value: no such baseline is reported.
There is another complication. Testing data were scaled separately, using a maximum drawn from the 2015 target year. This introduces information from the period being predicted into preprocessing, weakening the assessment as a simulation of forecasting an unknown future. No later, independent evaluation is reported after selecting the preferred network.
For readers assessing market forecasts, this study offers a bounded methodological lesson, not evidence about market growth or its causes. Before accepting an accuracy headline, check the scoring rule, the baseline and which information was available at prediction time.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Analysis of Artificial Neural Network Accuracy Using Backpropagation Algorithm In Predicting Process (Forecasting) Sandy Putra Siregar, Anjar Wanto
