Scholarlink

<p>Our aim is the development of effective methods in <i>out-of-sample</i> forecasting of cryptocurrency prices and price trend. With regard to Bitcoin price forecasting in particular, we present a hybrid ensemble that proves superior when tested against benchmark models (singular and ensemble). We demonstrate two sources of model superiority: One, our suggested hybrid ensemble models both linear functionality (with <i>ARIMA(p,d,q)</i>) and nonlinear functionality (with <i>NAR</i>) in a <i>BTC/USD</i> price time series data set; as opposed to only nonlinear modeling (which has been the conventional approach in a recent, documented trend of data driven hybrid modeling of cryptocurrency prices). Modeling both types of functionalities generates a broader set of independent information regarding target variable movement. This leads to a better overall pattern recognition in the time series data, which we demonstrate contributes to superior price forecast accuracy. Two, our suggested hybrid ensemble incorporates a novel, improved (and more theoretically sound) approach to data smoothing: Each data point in the time series is individually de-noised, rather than relying on the conventional approach of using a standardized variable to smooth the entire data series. Our smoothing approach leads to clearer indication of any nonlinear trend in the data, which we demonstrate further paves the way to superior forecast accuracy, when coupled with linear modeling in the hybrid. With regard to forecasting price trend, we find that a deep learning model, in the form of a <i>Gated Recurrent Unit (GRU)</i>, proves superior in generating <i>out-of-sample</i> forecasts of Bitcoin price trend. These forecasts closely align with realized price movement over the <i>out-of-sample</i> period of the present study. We demonstrate that a simulated, ex-ante Bitcoin investment in which long (short) positions are taken during <i>out-of-sample</i> intervals of forecasted upward (downward) price-trend outperforms a <i>buy and hold</i> strategy over the same, entire <i>out-of-sample</i> period.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Contemporary Approaches to Hybrid Forecasting

  • 期刊论文

  • 发表日期：

    2025年8月27日

查看全文
  • Ugur Sener,
  • Salvatore Joseph Terregrossa

摘要

Our aim is the development of effective methods in out-of-sample forecasting of cryptocurrency prices and price trend. With regard to Bitcoin price forecasting in particular, we present a hybrid ensemble that proves superior when tested against benchmark models (singular and ensemble). We demonstrate two sources of model superiority: One, our suggested hybrid ensemble models both linear functionality (with ARIMA(p,d,q)) and nonlinear functionality (with NAR) in a BTC/USD price time series data set; as opposed to only nonlinear modeling (which has been the conventional approach in a recent, documented trend of data driven hybrid modeling of cryptocurrency prices). Modeling both types of functionalities generates a broader set of independent information regarding target variable movement. This leads to a better overall pattern recognition in the time series data, which we demonstrate contributes to superior price forecast accuracy. Two, our suggested hybrid ensemble incorporates a novel, improved (and more theoretically sound) approach to data smoothing: Each data point in the time series is individually de-noised, rather than relying on the conventional approach of using a standardized variable to smooth the entire data series. Our smoothing approach leads to clearer indication of any nonlinear trend in the data, which we demonstrate further paves the way to superior forecast accuracy, when coupled with linear modeling in the hybrid. With regard to forecasting price trend, we find that a deep learning model, in the form of a Gated Recurrent Unit (GRU), proves superior in generating out-of-sample forecasts of Bitcoin price trend. These forecasts closely align with realized price movement over the out-of-sample period of the present study. We demonstrate that a simulated, ex-ante Bitcoin investment in which long (short) positions are taken during out-of-sample intervals of forecasted upward (downward) price-trend outperforms a buy and hold strategy over the same, entire out-of-sample period.