Mean Forecast Error Research Articles

Forecasting accuracy: a comparative study between artificial neural network and autoregressive model for streamflow

Estimating the reliability of potential prediction is very crucial as our life depended heavily on it. Thus, a simulation that concerned hydrological factors such as streamflow must be enhanced. In this study, Autoregressive (AR) and Artificial Neural Networks (ANN) were used. The forecasting result for each model was assessed by using various performance measurements such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Forecast Error (MFE) and Nash-Sutcliffe Model Efficiency Coefficient (CE). The experimental results showed the forecast performance of Durian Tunggal reservoir datasets by using ANN Model 7 with 7 hidden neurons has better forecast performance compared to AR (4). The ANN model has the smallest MAE (0.0116 m<sup>3</sup>/s), RMSE (0.0607 m<sup>3</sup>/s), MAPE (1.8214% m<sup>3</sup>/s), MFE (0.0058 m<sup>3</sup>/s) and largest CE (0.9957 m<sup>3</sup>/s) which show the capability of fitting to a nonlinear dataset. In conclusion, high predictive precision is an advantage as a proactive or precautionary measure that can be inferred in advance in order to avoid certain negative effects.

Forecasting the Daily 10.7 cm Solar Radio Flux Using an Autoregressive Model

As an important proxy of the solar extreme ultraviolet radiation from the upper chromosphere and lower corona, the 10.7 cm solar radio flux (F10.7) has a wide range of applications in models of the thermosphere and ionosphere. Forecasting F10.7 has already become a routine business in space weather services. In this study, we analyzed the predictive power of autoregressive (AR) models with orders $p=15$ – 1005, a training sample length $L=22$ years, and a running time window $w = 50$ days on the daily F10.7, during the last two solar cycles (Solar Cycles 23 and 24) at the forecast steps $n=1$ – 81 days. The main conclusions are as follows. (i) The mean forecast error ( $\overline{\delta }$ ) at the $n$ th day or over $N$ days is minimum at an optimal order $p_{\mathrm{o}}$ , which tends to increase as $n$ or $N$ increases. (ii) $\overline{\delta }$ is positively related to both $n$ and F10.7. The large error during the maximum period is the result of the large daily variation in F10.7, mainly due to the appearance and decay of active regions, especially the eruptions of solar flares. (iii) The solar cycle can be divided into six parts in the rising order of $\overline{\delta }$ : (a) closing part of the declining phase, (b) initial rising phase, (c) middle declining phase, (iv) closing rising phase, (v) middle rising phase, and (f) initial declining phase. (iv) The AR model at $p_{\mathrm{o}}$ is not inferior to other techniques. (v) $p_{\mathrm{o}}$ is uncorrelated to the autocorrelation coefficient, and (vi) $\overline{\delta }$ is minimum at a certain $L$ .

Forecast accuracy of COVID-19 spread dynamics in Peru

Objective: To analyze the forecast accuracy of Brown's exponential smoothing model to predict the spread of COVID-19 in Peru from March 6 to May 30, 2020. Materials and methods: A descriptive study based on a time series analysis conducted from March 6 to May 30, 2020 in Peru. The information on the number of positive cases of COVID-19 (155,671 people) was used. The prediction method was Brown's exponential smoothing model, which consists in carrying out two exponential smoothings from which the forecast is calculated: the time series values were used in the first smoothing, and the first attenuation series was used in the second one. Accuracy measures used in the research were: mean forecast error (MFE), mean squared error (MSE), mean absolute deviation (MAD) and mean absolute percentage error (MAPE). The coefficient of determination (R2) was used to establish if the data fits the evaluated model. Results: MFE was 156.7, MSE was 506461.3, MAD was 450.6 and MAPE was 9.03 %. R2 accounted for 0.8078. Conclusions: Accuracy error or MAPE was 9.03 % and R2 was 0.8078, which indicates that the data fits by 80.78 % to the evaluated model.

The Use of Gridded Model Output Statistics (GMOS) in Energy Forecasting of a Solar Car

For many years, primary weather forecasting services (Global Forecast System (GFS) and the European Centre for Medium-Range Weather Forecasts (ECMWF)) have been made available to the public through global Numerical Weather Prediction (NWP) models estimating a multitude of general weather variables in a variety of resolutions. Secondary services such as weather experts Meteomatics AG use data and improve the forecasts through various methods. They tailor for the specific needs of customers in the wind and solar power generation sector as well as data scientists, analysts, and meteorologists in all areas of business. These auxiliary services have improved performance and provide reliable data. However, this work extended these auxiliary services to so-called tertiary services in which the weather forecasts were further conditioned for the very niche application environment of mobile solar technology in solar car energy management. The Gridded Model Output Statistics (GMOS) Global Horizontal Irradiance (GHI) model developed in this work utilizes historical data from various ground station locations in South Africa to reduce the mean forecast error of the GHI component. An average Root Mean Square Error (RMSE) improvement of 11.28% was shown across all locations and weather conditions. It was also shown how the incorporation of the GMOS model could have increased the accuracy in regard to the State of Charge (SoC) energy simulation of a solar car during the Sasol Solar Challenge 2018 and the possible range benefits thereof.

The Impact of Dropsonde and Extra Radiosonde Observations during NAWDEX in Autumn 2016

Abstract Dropsonde observations from three research aircraft in the North Atlantic region, as well as several hundred additionally launched radiosondes over Canada and Europe, were collected during the international North Atlantic Waveguide and Downstream Impact Experiment (NAWDEX) in autumn 2016. In addition, over 1000 dropsondes were deployed during NOAA’s Sensing Hazards with Operational Unmanned Technology (SHOUT) and Reconnaissance missions in the west Atlantic basin, supplementing the conventional observing network for several intensive observation periods. This unique dataset was assimilated within the framework of cycled data denial experiments for a 1-month period performed with the global model of the ECMWF. Results show a slightly reduced mean forecast error (1%–3%) over the northern Atlantic and Europe by assimilating these additional observations, with the most prominent error reductions being linked to Tropical Storm Karl, Cyclones Matthew and Nicole, and their subsequent interaction with the midlatitude waveguide. The evaluation of Forecast Sensitivity to Observation Impact (FSOI) indicates that the largest impact is due to dropsondes near tropical storms and cyclones, followed by dropsondes over the northern Atlantic and additional Canadian radiosondes. Additional radiosondes over Europe showed a comparatively small beneficial impact.

Uniform Predictive Inference for Factor Models with Instrumental and Idiosyncratic Betas

This paper investigates the effect of characteristic-based time-varying factor beta on the diffusion-index type forecast. Specifically, the factor beta includes two distinct components: the instrumental beta'' is a function of some observed stable variables, while the beta'' captures the more volatile residual movements. To estimate these time-varying beta's, we propose to apply the projected PCA (P-PCA), which takes advantage of firm characteristics, on high-frequency return data. We show that various components of the conditional mean forecast error are all asymptotically normal. Of particular note is that the time-varying idiosyncratic beta has nontrivial effects on the first-order asymptotic of the conditional mean forecast. A Monte Carlo exercise demonstrates fairly good finite sample properties of the P-PCA estimators and shows that the P-PCA method largely outperforms the traditional PCA method for factor-augmented prediction. In our empirical experiments of volatility forecasting, we find that the predictive model estimated by the P-PCA method achieves preferable performance for a wide variety of target assets. We provide evidence that the forecast uncertainty arose from the time variation in factor beta magnifies during volatile periods. In such cases, exploiting the information in stable characteristics plays an even more crucial role in producing effective diffusion-index forecasts.

How accurate are A-REIT IPO dividend forecasts?

Purpose The purpose of this paper is to investigate the accuracy of Australian Real Estate Investment Trust (A-REIT) initial public offering (IPO) dividend forecasts between 1994 and 2016. Design/methodology/approach This study compares the dividend forecasts of A-REIT IPOs for the first dividend forecast period in the prospectus, with the actual dividend declared for that forecast period. As well as simple descriptive summary measures, this study also employs an exact logistic regression approach to examine the factors that might influence the IPOs achieving or exceeding the dividend forecast. Findings The study identifies that the dividends declared, on average, were greater than the dividend forecast and that more than nine out of ten of the IPOs listed after 1999 achieved or exceeded their prospectus forecast. In addition the authors observe positive mean forecast errors, suggesting dividend forecasts in A-REIT IPOs, are cautiously biased. This is in contrast to the industrial company data reported in Brown et al. (2000) which suggest dividend forecasts are optimistically biased. The study also finds the A-REIT IPOs that did not forecast a dividend, generally did not pay a dividend. Practical implications The results will inform dividend seeking institutional and retail investors of the investment opportunities in A-REIT IPOs. Originality/value This paper adds to the discussion of the relative predictability of dividends of A-REIT IPOs compared to industrial company IPOs.

The accuracy of earnings forecast in IPO prospectuses: Evidence from Indonesia

The financial information that attracts the attention of most investors is the earnings forecast. Because investors are more interested in the company's prospects in the future than historical information. But in reality earnings forecast is less accurate, while the factors that are suspected to influence give inconsistent findings. This study aims to determine the effect of the forecast horizon, underwriter reputation, auditor reputation, company size, company age, leverage ratio, industry type and IPO market on the accuracy of earnings forecast. The population of all companies that IPO in Indonesia Stock Exchange period 2015-2018, were selected using the purposive sampling method, obtained 107 samples. Secondary data were obtained from the company's prospectus and annual report. The analytical method used is multiple regression. The accuracy level of earnings forecasts is measured using forecast error numbers. Mean forecast error 28.86 percent is reported over the entire sample period. Then only underwriter reputation and company size have a significant positive effect on the accuracy of earnings forecasts. And the hot IPO market has a negative effect on the accuracy of earnings forecasts. The implications of research show that underwriter reputation, company size, and IPO market became the important indicators in determining investment strategies in IPO shares. JEL Classification: D82, G14, G17, R53 DOI: https://doi.org/10.26905/jkdp.v23i4.3509

ПОВЫШЕНИЕ ТОЧНОСТИ ПРОГНОЗИРОВАНИЯ БИРЖЕВЫХ ЦЕН НИКЕЛЯ НА ОСНОВЕ КОМПЛЕКСНОГО ПРИМЕНЕНИЯ РЯДА МОДЕЛЕЙ

The paper focuses on the fact that Russia is one of the main exporters of ferrous and non-ferrous metals on the world market. A large sector of producers and traders with this raw material is represented in the country. For successful trading it is necessary to know the prices in the future, when the contract is concluded for the export delivery of a consignment of goods, the parties often fix the price at the time of signing the treaty. By the time of cargo shipment the price may suddenly change, so that the transaction would be disadvantageous for the exporter. Today, companies are actively using forecasting and, in the case of the global market, they are guided by global stock quotes. The prices on base metals correlate well with each other, for simplicity, and forecasting nickel price quotes is more available to take as an example. It has been stated that from the time of making the contract to the time of delivery there pass 14 days, which requires forecasting 14 steps ahead (long warning period). This fact causes the main difficulty in achieving the required accuracy. The prediction accuracy is suggested by using simple and weighted averaging of the predicted values of several models. The results of using several models of common statistical forecasting and developed in the course of the research are presented. The mean forecasting error of nickel price quotations for the period from March, 2015 to March, 2016 for 14 days ahead for the best of the given values made 6.48%, for the worst value - 14.88%. The simple averaging value has shown a prediction error of 6.01%, and the weighted value - 3.58%. It has been inferred that using this technique improves the accuracy of the available forecasting models in practice.

Imputation of Missing Values in Economic and Financial Time Series Data Using Five Principal Component Analysis (PCA) Approaches

This study assessed five approaches for imputing missing values. The evaluated methods include Singular Value Decomposition Imputation (svdPCA), Bayesian imputation (bPCA), Probabilistic imputation (pPCA), Non-Linear Iterative Partial Least squares imputation (nipalsPCA) and Local Least Squares imputation (llsPCA). A 5%, 10%, 15% and 20% missing data were created under a missing completely at random (MCAR) assumption using five (5) variables (Net Foreign Assets (NFA), Credit to Core Private Sector (CCP), Reserve Money (RM), Narrow Money (M1), Private Sector Demand Deposits (PSDD) from Nigeria quarterly monetary aggregate dataset from 1981 to 2019 using R-software. The data were collected from the Central Bank of Nigeria statistical bulletin. The five imputation methods were used to estimate the artificially generated missing values. The performances of the PCA imputation approaches were evaluated based on the Mean Forecast Error (MFE), Root Mean Squared Error (RMSE) and Normalized Root Mean Squared Error (NRMSE) criteria. The result suggests that the bPCA, llsPCA and pPCA methods performed better than other imputation methods with the bPCA being the more appropriate method and llsPCA, the best method as it appears to be more stable than others in terms of the proportion of missingness.

Impact of Multiple Altimeter Data and Mean Dynamic Topography in a Global Analysis and Forecasting System

AbstractSatellite altimetry is one of the main sources of information used to constrain global ocean analysis and forecasting systems. In addition to in situ vertical temperature and salinity profiles and sea surface temperature (SST) data, sea level anomalies (SLA) from multiple altimeters are assimilated through the knowledge of a surface reference, the mean dynamic topography (MDT). The quality of analyses and forecasts mainly depends on the availability of SLA observations and on the accuracy of the MDT. A series of observing system evaluations (OSEs) were conducted to assess the relative importance of the number of assimilated altimeters and the accuracy of the MDT in a Mercator Ocean global 1/4° ocean data assimilation system. Dedicated tools were used to quantify impacts on analyzed and forecast sea surface height and temperature/salinity in deeper layers. The study shows that a constellation of four altimeters associated with a precise MDT is required to adequately describe and predict upper-ocean circulation in a global 1/4° ocean data assimilation system. Compared to a one-altimeter configuration, a four-altimeter configuration reduces the mean forecast error by about 30%, but the reduction can reach more than 80% in western boundary current (WBC) regions. The use of the most recent MDT updates improves the accuracy of analyses and forecasts to the same extent as assimilating a fourth altimeter.

A comparison of GSTAR-SUR models and a hybrid GSTAR-SUR/neural network model on residuals of precipitation forecasting

The Generalized Space Time Autoregressive-Seemingly Unrelated Regression (GSTAR-SUR) model is often used to forecast data that have time and location components. At the current time, precipitation is difficult to predict because it has patterns and characteristics that are hard to identify. This phenomenon is referred to as a non-linear phenomenon. One model that considers non-linearity is the neural network. The GSTAR-SUR model is a linear model, so at the current time is needed non-linear model on precipitation forecasting. This current research compares the precipitation forecast results from the GSTAR-SUR model and a hybrid GSTAR-SUR-with-neural-network approach in residuals. The data used in this research are the precipitation records for four locations in West Java for the years 2005 to 2015. The precipitation data represent 10-day-long observations. Precipitation in the four locations will be modeled using two approaches, i.e., GSTAR-SUR and GSTAR-SUR-NN. In the GSTAR-SUR-NN model, the residuals of the GSTAR-SUR model will be the basis of a neural network model. In this case, the GSTAR-SUR-NN model resulted in forecast data that are closer to the observed values than those from the GSTAR-SUR model. The mean forecasting error of the GSTAR-SUR-NN model was 3.8917 mm, while that of the GSTAR-SUR model was 4.3295 mm.

The Relationship between Deterministic and Ensemble Mean Forecast Errors Revealed by Global and Local Attractor Radii

It has been demonstrated that ensemble mean forecasts, in the context of the sample mean, have higher forecasting skill than deterministic (or single) forecasts. However, few studies have focused on quantifying the relationship between their forecast errors, especially in individual prediction cases. Clarification of the characteristics of deterministic and ensemble mean forecasts from the perspective of attractors of dynamical systems has also rarely been involved. In this paper, two attractor statistics—namely, the global and local attractor radii (GAR and LAR, respectively)—are applied to reveal the relationship between deterministic and ensemble mean forecast errors. The practical forecast experiments are implemented in a perfect model scenario with the Lorenz96 model as the numerical results for verification. The sample mean errors of deterministic and ensemble mean forecasts can be expressed by GAR and LAR, respectively, and their ratio is found to approach $$\sqrt 2 $$ with lead time. Meanwhile, the LAR can provide the expected ratio of the ensemble mean and deterministic forecast errors in individual cases.

Cross-Validation Based Forecasting Method: A Machine Learning Approach

Our paper aims to evaluate two novel methods on selecting the forecasting model or its combination based on a Machine Learning approach. The methods are based on the selection of the best model, or combination of models, by cross-validation technique, from a set of possible models. The first one is based on the seminal paper of Granger-Bates (1969) but weights are estimated by a process of cross-validation applied on the training set. The second one selects the model with the forecasting performance in the process described above, which we called CvML (Cross-Validation Machine Learning Method). The following models are used: exponential smoothing, SARIMA, artificial neural networks and Threshold autoregression (TAR). Model specification is chosen by R packages: forecast and TSA. Both methods -- CvML and MGB -- are applied to these models to generate forecasts from one up to twelve periods ahead. Frequency of data is monthly. We run the forecasts exercise to the following to monthly series of Industrial Product Indices for seven countries: Canada, Brazil, Belgium, Germany, Portugal, UK and USA. The data was collected at OECD data, with 504 observations. We choose Average Forecast Combination, Granger Bates Method, MCS model, Naive and Seasonal Naive Model as benchmarks. Our results suggest that MGB did not performed well. However, CvML had a lower mean absolute error for most of countries and forecast horizons, particularly at longer horizons, surpassing all the proposed benchmarks. Similar results hold for absolute mean forecast error.

Application of Support Vector Regression for Modeling Low Flow Time Series

Hydrologic time series modeling using historical records plays a crucial role in forecasting different hydrological processes. The objective of this study is to analyze the suitability of Support Vector Regression (SVR) for modeling monthly low flows time series for three stations in Mahanadi river basin, India. The ‘low flow’ threshold was taken as the Q75 discharge, i.e., the flow is equal to or surpassed for the duration of 75% of the observation period which was obtained from the daily discharge data. The potential applicability of SVR model is assessed with two different framework models (ANN-ELM, GPR) based on various statistical measures (r2, RMSE, MAE, Nash-Sutcliffe coefficient, objective function (OBJ), Scatter Index (SI) and BIAS). The model selection was based on lowest OBJ value for each station amongst three models (SVR, ANN-ELM, GPR). The SVR model was trained using the Radial Basis Function (RBF). Using the same inputs, the other two models (ANN-ELM and GPR) was also tested. From results, among all the stations, the SVR outperformed GPR and ANN-ELM with lowest OBJ value for the three stations a (1.378, 1.202, 1.570). In addition, the accuracy of the three models were checked using mean forecasting error which were (0.474, 0.421, 0.509) for SVR, (0.507, 0.489 0.500) for GPR and (0.564, 0.603, 0.772) ANN-ELM for the three stations. The results confirm that SVR can be used satisfactorily for modeling monthly low flows in the Mahanadi river basin, India. Hence, the SVR model could be employed as a new reliable and accurate data intelligent approach for predicting the ‘low flow’ (Q75 discharge) based on precedent data in water resources and its allied field.

Development of NWP Based Objective Cyclone Prediction System (CPS) for North Indian Ocean Tropical Cyclones — Evaluation of Performance

ABSTRACT An objective NWP based Cyclone Prediction System (CPS) has been developed and implemented at IMD for operational cyclone forecasting over the north Indian Ocean (NIO). The five forecast components of CPS are (a) Genesis Potential Parameter (GPP), (b) Multi-Model Ensemble (MME) technique for track prediction, (c) Statistical Cyclone Intensity Prediction (SCIP), (d) rapid intensification and (e) decay model to forecast intensity after the landfall. Verification shows GPP had higher probability of detection (0.98) and lower false alarm ratio (0.27) with higher critical success index (0.72). Mean track error of MME ranged from 74 km at 12 h to 200 km at 72 h and reduced by 27% to 52% for 36 h to 72 h forecast during 2009-2013. The mean forecast errors of landfall position ranged from 56 km at 24 h to 137 km at 72 h and landfall time error ranged from 3.6 h at 24 h to 6.1 h at 72 h. Mean intensity errors of SCIP ranged from 5.4 kt at 12 h to 16.9 kt at 72 h. The probabilistic rapid intensification forecast was skillful compared to climatology. The 6-hourly decaying intensity (after landfall) errors ranged from 3 kt to 4.9 kt. Results demonstrate the potential of CPS for operational cyclone forecast over the NIO.

Point forecasting of intraday volume using Bayesian autoregressive conditional volume models

AbstractIn this paper, we apply Bayesian inference to model and forecast intraday trading volume, using autoregressive conditional volume (ACV) models, and we evaluate the quality of volume point forecasts. In the empirical application, we focus on the analysis of both in‐ and out‐of‐sample performance of Bayesian ACV models estimated for 2‐minute trading volume data for stocks quoted on the Warsaw Stock Exchange in Poland. We calculate two types of point forecasts, using either expected values or medians of predictive distributions. We conclude that, in general, all considered models generate significantly biased forecasts. We also observe that the considered models significantly outperform such benchmarks as the naïve or rolling means forecasts. Moreover, in terms of root mean squared forecast errors, point predictions obtained within the ACV model with exponential distribution emerge superior compared to those calculated in structures with more general innovation distributions, although in many cases this characteristic turns out to be statistically insignificant. On the other hand, when comparing mean absolute forecast errors, the median forecasts obtained within the ACV models with Burr and generalized gamma distribution are found to be statistically better than other forecasts.

A Study on an Automatical BKLS Measurement By Programming Technology

This study focuses on presenting the IT program module provided by BKLS measure in order to solve the problem of capital cost due to information asymmetry of external investors and corporate executives. Barron at al(1998) set up a BKLS measure to guide the market by intermediate analysts. The BKLS measure was measured by using the changes in the analyst forecast dispersion and analyst mean forecast error squared. This study suggests a model of the algorithm that the BKLS measure can be provided to all investors immediately by IT program in order to deliver the meaningful value in the domestic capital market as measured. This is a method of generating and analyzing real-time or non-real-time prediction models by transferring the predicted estimates delivered to the Big Data Log Analysis System through the statistical DB to the statistical forecasting engine. Because BKLS measure is not carried out in a concrete method, it is practically very difficult to estimate the BKLS measure. It is expected that the BKLS measure of Barron at al(1998) introduced in this study and the model of IT module provided in real time will be the starting point for the follow-up study for the introduction and realization of IT technology in the future..

Thou should not panic! Let calmness fight the Crocodile Bite

This paper proposes a rule of a thumb, named as crocodile rule for Pakistan Stock Exchange, which is an emerging stock market. According to this rule when a crocodile attacks the only recourse is to stop the struggle to safeguard. In this way, the crocodile will assume the person is dead and it will stop the attack. If applied to stock markets, the rule portrays that when performance falls below a threshold level the investors must not pull back the funds and shall continue their investments, as there is significant chance of recovery. The rule is tested in two steps. Firstly, we will check by forecasting the probability. Secondly, we will use these probabilities to test the rule of thumb. For empirical analysis, data is taken on monthly basis from 2000 to 2017. The results for the first step show that mean forecasting error for all categories of stocks is close to zero. Furthermore, the rule is valid for all three categories of stocks; risky, size and value on Pakistan Stock Exchange. The results from 2000–2017 show, that on average there is 74% chance of recoverywith minimum and maximum recovery chances of 66% and 83% respectively. Where 83% chance is very optimistic, 66% chance seems quite low, but investors must remember no alarming situation can be dealt with fright and thus, thou should not panic! Let calmness fight the crocodile bite.

A support vector machine for model selection in demand forecasting applications

Abstract Time series forecasting has been an active research area for decades, receiving considerable attention from very different domains, such as econometrics, statistics, engineering, mathematics, medicine and social sciences. Moreover, with the emergence of the big data era, the automatic identification with the appropriate techniques remains an intermediate compulsory stage of any big data implementation with predictive analytics purposes. Extensive research on model selection and combination has revealed the benefits of such techniques in terms of forecast accuracy and reliability. Several criteria for model selection have been proposed and used for decades with very good results. Akaike information criterion and Schwarz Bayesian criterion are two of the most popular criteria. However, research on the combination of several criteria along with other sources of information in a unified methodology remains scarce. This study proposes a new model selection approach that combines different criteria using a support vector machine (SVM). Given a set of candidate models, rather than considering any individual criterion, an SVM is trained at each forecasting origin to select the best model. This methodology will be particularly interesting for scenarios with highly volatile demand because it allows changing the model when it does not fit the data sufficiently well, thereby reducing the risk of misusing modeling techniques in the automatic processing of large datasets. The effects of the proposed approach are empirically explored using a set of representative forecasting methods and a dataset of 229 weekly demand series from a leading household and personal care manufacturer in the UK. Our findings suggest that the proposed approach results in more robust predictions with lower mean forecasting error and biases than base forecasts.