A Dual Network Solution (DNS) for Lag-Free Time Series Forecasting Subhrajit Samanta ERI@N-IGS NTU Singapore sama0021@e.ntu.edu.sg Mahardhika Pratama SCSE NTU Singapore MPRATAMA@ntu.edu.sg Suresh Sundaram Aerospace Engineering IISC Bengaluru, India sureshsundaram@iisc.ac.in Narasimalu Srikanth ERI@N NTU Singapore nsrikanth@ntu.edu.sg Abstract—When it comes to time series forecasting, lag in the predicted sequence can be a predominant issue. Unfortunately, this is often overlooked in most of the time series literature as this does not contribute to a high prediction error (i.e. MSE). However, it leads to a rather poor forecast in terms of movement prediction in time series. In this article, we tackle this basic problem with a novel trend driven mechanism. Trend, defined as the inherent pattern of the data, is extracted here and utilized next to perform a lag-free forecasting. We propose a generic and light Dual Network Solution (DNS), where the first network predicts the trend and the second network utilizes that predicted trend along with its historical information to capture the dynamical behavior of the time series efficiently. DNS exhibits a substantially improved (≈ 10% better) performance compared to more complex and resource-intensive state-of-the- art algorithms in large scale regression problems. Apart from the traditional Mean Squared Error (MSE), we also propose a new Movement Prediction Metric or MPM (for detection of lag in time series) as a new complementary performance metric to evaluate the efficacy of DNS better. Index Terms—Lag in time series prediction, Dual network solution, Movement prediction, Lag free time series prediction I. I NTRODUCTION Time series modelling and forecasting are one of the most popular areas of research in the machine learning and data science community for its wide applicability in various domains across different sectors such as energy demand prediction, weather forecasting, financial prediction, etc. This ever-growing field of study has seen significant improvement over the past decades. Starting from the statistical methods of regression analysis to recent advents in machine learning algorithms have propelled this research to a new high. However, across most of this literature, a basic problem of lag difference (in the predicted sequence) persists. Next, we discuss the problem elaborately to highlight our motivation. A. Problem Description : Lag in Time Series A time series model utilizing historical data alone can detect a change in the trend only after the historical data also experiences the same change. Naturally, the change prediction is late which leads to the problem of lag. To explain this problem further we take an example case: wind speed 1 forecasting. The 1 http://mesonet.agron.iastate.edu/request/awos/1min.php prediction task is performed here with a Multi-Layer Perceptron network or MLP (hidden node with sigmoid activations in the single hidden layer) with only past values of the target sequence as inputs and the actual vs prediction plot is provided in figure 1. From time instance 25 to 31 in figure 1b the trend is upward which leads the MLP to predict an upward trend at point 31 (marked in the figure 1b) in spite of an actual downward trend. This results in a predicted sequence with a distinct lag as seen in the figure 1a and 1b. 0 100 200 300 400 500 600 Data Points -1 -0.5 0 0.5 1 Normalized Wind Speed Actual speed Predicted Speed Lag in predicted sequence (a) Actual vs prediction plot 25 26 27 28 29 30 31 32 33 34 35 Data Points -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 Normalized Wind Speed Data points Actual Trend Predicted Trend Actual Trend Changes here Predicted Trend Changes here Lag (b) Zoomed in to show Lag Fig. 1: Demonstration of Lag in Time Series Forecasting The lag in the predicted sequence results in a ’close-by, past value’ prediction hence it does not contribute to a high prediction error such as MSE. Besides, MSE does not take into account the movement direction of the prediction (as the squared error ignores its sign), therefore it is unable to detect lag. For instance, the wind forecast in figure 1a looks like a good forecast however suffers from the lag problem. Sometimes due to over-fitting of the historical values, the result can be precisely a curve that mimics the real values almost 978-1-7281-6926-2/20/$31.00 ©2020 IEEE