Publication: Yapay sinir ağları ve gri model ile döviz kuru tahmini
Abstract
Geçmiş yıllarda, zaman serisi tahmin problemleri için çeşitli istatistiksel yöntemler kullanıldı. Günümüzde artık geleneksel istatistiksel tahminler yetersiz kalmaktadır. Bunun en önemli nedeni veri boyutlarının artmasıdır. Buna bağlı olarak zaman serisi verilerininözellikleri ve ilişkilerinin belirlenmesi zorlaşmakta ve bu karmaşıklığı çözme kabiliyetine sahip Derin Yapay Sinir Ağları önerilmektedir. LSTM katmanları ve algoritması, derin sinir ağlarında yeni bir yöntem ve zaman serileri tahmininde profesyonel çözüm olmaktadır. Tekrarlayan Sinir Ağları (Recurrent neural network - RNN) nın öğrenememe sorunu, Hochreiter ve Schmidhuber tarafından detaylı araştırılarak yeni bir model olarak LSTM geliştirilmiş ve 1997 yılında bir makale ile tanıtılmıştır.Bu tez çalışmamda, döviz kuru tahmini için Gri tahmin modelleri de dikkate alınmıştır. Alternatif bir yöntem olarak, belirisizliklerin netleştirilmesinde uzun yıllardır kullanılan gri tahmin modelleri, önceden herhangi bir ön bilgiye ihtiyaç duymadan sınırlı veri ile başarılı tahminler gerçekleştirebilmektedir. Gri tahmin, zaman serisi ve sebep-sonuç ilişkisine dayalı çeşitli tahmin modellerini içermektedir. Çalışma kapsamında, zaman serisi modellerinden GM(1,1) gri tahmin modeli ele alınmıştır. Döviz kuru alış değerlerinin tahmini için T.C. Merkez Bankası 10 Temmuz 2019 ve 7 Aralık 2020 tarihleri arasındaki döviz kurları kullanılmıştır. CSV formatindaki döviz kuru verileri, YSA (LSTM) ve Gri model tahmin modellerine göre değerlendirilip oluşan sonuçlar sayısal ve görsel olarak karşılaştırılmıştır. Bu işlemler içinde Python programlama dili kullaılmış ve ilgili kodlar yazılmıştır.
In past years, various statistical methods have been used for time series prediction problems. Currently, the traditional statistical estimate is no longer sufficient. The most important reason for this is the increase in data sizes. Because of this, determining the properties and relationships of time series data becomes difficult, and Deep Artificial Neural Networks with the ability to solve this complexity are proposed.The LSTM layers and algorithm is a new method in deep neural networks and a professional solution in time series estimation. The problem of the inability to learn repetitive neural networks (RNN) was investigated in detail by Hochreiter and Schmidhuber, and LSTM was developed as a new model and introduced in 1997 with a paper.In this thesis study, Grey forecast models for exchange rate estimation were also taken into account. As an alternative method, Gray forecasting models, which have been used for many years to clarify uncertainties, can perform successful estimates with limited data without the need for any prior knowledge. Gray forecasting includes several forecasting models based on time series and cause-and-effect relationships. In the scope of the study, the GM(1,1) Gray prediction model from time series models was considered.For an estimate of exchange rate buying values, t.C. Dec July 10, 2019 and December 7, 2020 exchange rates were used by the central bank. Exchange rate data in CSV format were evaluated according to YSA (LSTM) and Gray model forecast models and the results were compared numerically and visually. In these operations, the Python programming language is used and the corresponding code is written.
In past years, various statistical methods have been used for time series prediction problems. Currently, the traditional statistical estimate is no longer sufficient. The most important reason for this is the increase in data sizes. Because of this, determining the properties and relationships of time series data becomes difficult, and Deep Artificial Neural Networks with the ability to solve this complexity are proposed.The LSTM layers and algorithm is a new method in deep neural networks and a professional solution in time series estimation. The problem of the inability to learn repetitive neural networks (RNN) was investigated in detail by Hochreiter and Schmidhuber, and LSTM was developed as a new model and introduced in 1997 with a paper.In this thesis study, Grey forecast models for exchange rate estimation were also taken into account. As an alternative method, Gray forecasting models, which have been used for many years to clarify uncertainties, can perform successful estimates with limited data without the need for any prior knowledge. Gray forecasting includes several forecasting models based on time series and cause-and-effect relationships. In the scope of the study, the GM(1,1) Gray prediction model from time series models was considered.For an estimate of exchange rate buying values, t.C. Dec July 10, 2019 and December 7, 2020 exchange rates were used by the central bank. Exchange rate data in CSV format were evaluated according to YSA (LSTM) and Gray model forecast models and the results were compared numerically and visually. In these operations, the Python programming language is used and the corresponding code is written.