Search Engine Data Research Articles

Deep learning models for hepatitis E incidence prediction leveraging Baidu index.

Infectious diseases are major medical and social challenges of the 21st century. Accurately predicting incidence is of great significance for public health organizations to prevent the spread of diseases. Internet search engine data, like Baidu search index, may be useful for analyzing epidemics and improving prediction. We collected data on hepatitis E incidence and cases in Shandong province from January 2009 to December 2022 are extracted. Baidu index is available from January 2009 to December 2022. Employing Pearson correlation analysis, we validated the relationship between the Baidu index and hepatitis E incidence. We utilized various LSTM architectures, including LSTM, stacked LSTM, attention-based LSTM, and attention-based stacked LSTM, to forecast hepatitis E incidence both with and without incorporating the Baidu index. Meanwhile, we introduce KAN to LSTM models for improving nonlinear learning capability. The performance of models are evaluated by three standard quality metrics, including root mean square error(RMSE), mean absolute percentage error(MAPE) and mean absolute error(MAE). Adjusting for the Baidu index altered the correlation between hepatitis E incidence and the Baidu index from -0.1654 to 0.1733. Without Baidu index, we obtained 17.04±0.13%, 17.19±0.57%, in terms of MAPE, by LSTM and attention based stacked LSTM, respectively. With the Baidu index, we obtained 15.36±0.16%, 15.15±0.07%, in term of MAPE, by the same methods. The prediction accuracy increased by 2%. The methods with KAN can improve the performance by 0.3%. More detailed results are shown in results section of this paper. Our experiments reveal a weak correlation and similar trends between the Baidu index and hepatitis E incidence. Baidu index proves to be valuable for predicting hepatitis E incidence. Furthermore, stack layers and KAN can also improve the representational ability of LSTM models.

Utilizing textual data from online reviews for daily tourism demand forecasting: A deep learning approach leveraging word embedding techniques

Accurately estimating daily tourism volumes is crucial for optimizing operational strategies and enhancing visitor experiences at tourist destinations. In this study, we leverage historical tourism volume data, search engine data, and online review data (including textual review contents) to forecast daily tourism demand. We develop a deep learning framework that includes a feature selection module to select search engine indices, a word embedding module to transform review texts into numerical predictors, and an extreme learning machine (ELM) enhanced with the whale optimization algorithm (WOA) for predictive modeling. Based on different word embedding techniques, we investigated two specific forecasting methods: one based on term frequency-inverse document frequency (TF-IDF) and Transformer, and the other based on the standard pre-trained bidirectional encoder representations from Transformers (BERT). These two methods enhance predictive accuracy while ensuring fast training and inference speeds, making it suitable for the high-frequency daily tourism forecasting task. Experimental results demonstrate that the two methods significantly outperform traditional approaches that rely solely on numerical data sources, achieving lower prediction errors and faster inference speeds compared to established benchmark methods, highlighting their potential to contribute to advancements in tourism prediction methodologies.

Mapping the Characteristics of Respiratory Infectious Disease Epidemics in China Based on the Baidu Index from November 2022 to January 2023.

Infectious diseases pose a significant global health and economic burden, underscoring the critical need for precise predictive models. The Baidu index provides enhanced real-time surveillance capabilities that augment traditional systems. Baidu search engine data on the keyword "fever" were extracted from 255 cities in China from November 2022 to January 2023. Onset and peak dates for influenza epidemics were identified by testing various criteria that combined thresholds and consecutive days. The most effective scenario for indicating epidemic commencement involved a 90th percentile threshold exceeded for seven consecutive days, minimizing false starts. Peak detection was optimized using a 7-day moving average, balancing stability and precision. The use of internet search data, such as the Baidu index, significantly improves the timeliness and accuracy of disease surveillance models. This innovative approach supports faster public health interventions and demonstrates its potential for enhancing epidemic monitoring and response efforts.

Google Internet searches related to inflammatory arthritis: An observational study using Google Trends data.

The Internet has transformed how patients access health information. We examined Google search engine data to understand which aspects of health are most often searched for in combination with inflammatory arthritis (IA). Using Google Trends data (2011-2022) we determined the relative popularity of searches for 'patient symptoms' (pain, fatigue, stiffness, mood, work) and 'treat-to-target' (disease-modifying drugs, steroids, swelling, inflammation) health domains made with rheumatoid arthritis (RA), psoriatic arthritis (PsA), and axial spondyloarthritis (AxSpA) in the UK/USA. Google Trends normalises searches by popularity over time and region, generating 0-100 scale relative search volumes (RSV; 100 represents the time-point with most searches). Up to five search term combinations can be compared. In all IA forms, pain was the most popular patient symptom domain. UK/USA searches for pain gave mean RSVs of 58/79, 34/51, and 39/63 with RA, PsA, and AxSpA; mean UK/USA RSVs for other patient symptom domains ranged 2-7/2-8. Methotrexate was the most popular treat-to-target search term with RA/PsA in the UK (mean 28/21) and USA (mean 63/33). For AxSpA, inflammation was most popular (mean UK/USA 9/34). Searches for pain were substantially more popular than searches for methotrexate in RA and PsA, and inflammation in AxSpA. Searches increased over time. Pain is the most popular search term used with IA in Google searches in the UK/USA, supporting surveys/qualitative studies highlighting the importance of improving pain to patients with IA. Routine pain assessments should be embedded within treat-to-target strategies to ensure patient perspectives are considered.

Exploring Appropriate Search Engine Data for Interval Tourism Demand Forecasting Responding a Public Crisis in Macao: A Combined Bayesian Model

Public crises can bring unprecedented damage to the tourism industry and challenges to tourism demand forecasting, which is essential for crisis management and sustainable development. Existing studies mainly focused on point forecasts, but point forecasts may not be enough for the uncertain environments of public crises. This study proposes a combined Bayesian interval tourism demand forecasting model based on a forgetting curve. Moreover, considering tourists’ travel plans may be adjusted due to changing crisis situations, the choice of search engine data for forecasting tourism demand is investigated and incorporated into the proposed model to yield reliable results. Through an empirical study, this study figures out that the Baidu Index had better tourism predictive capabilities before the public crisis, whereas the Google Index effectively captured short-term fluctuations of tourism demand within the crisis period. The results also indicate that integrating both Baidu and Google Index data obtains the best prediction performance after the crisis outbreak. Our main contribution is that this study can generate flexible forecasting results in the interval form, which can effectively handle uncertainties in practice and formulate control measures for practitioners. Another novelty is successfully discovering how to select appropriate search engine data to improve the performance of tourism demand forecasts across different stages of a public crisis, thus benefiting daily operations and crisis management in the tourism sector.

Assessing the reactions of tourist markets to reinstated travel restrictions in the destination during the post-COVID-19 phase

This study, leveraging search engine data, investigates the dynamics of China's domestic tourism markets in response to the August 2022 epidemic outbreak in Xinjiang. It focuses on understanding the reaction mechanisms of tourist-origin markets during destination crises in the post-pandemic phase. Notably, the research identifies a continuous rise in the potential tourism demand from tourist origin cities, despite the challenges posed by the epidemic. Further analysis uncovers a regional disparity in the growth of tourism demand, primarily influenced by the economic stratification of origin markets. Additionally, the study examines key tourism attractions such as Duku Road, highlighting its resilient competitive system, which consists of distinctive tourism experiences, economically robust tourist origins, diverse tourist markets, and spatial pattern stability driven by economic factors in source cities, illustrating an adaptive response to external challenges such as crises. The findings provide new insights into the dynamics of tourism demand, offering a foundation for developing strategies to bolster destination resilience and competitiveness in times of health crises.

Evolution characteristics and influencing factors of information network in Guangdong-Hong Kong-Macao Greater Bay Area.

In the context of the digital information era, the impact of "The Internet Plus," "Big Data," and other technologies on urban social development has been far beyond any preceding era, under the influence of information technology, urban agglomeration space exhibits a new layout. Based on the search engine data of eleven cities in the Guangdong-Hong Kong-Macao Greater Bay Area from 2012 to 2021, this research constructs the inter-city information network strength linkage matrix to examine the evolution characteristics of city network structure and its driving causes. The results reveal that (1) the overall information linkage strength exhibits a pattern of steadily growing the radiating effect from the leading cities of Guangdong, Shenzhen, and Hong Kong to the surrounding cities, and a closer and more balanced information linkage network is gradually built. (2) Guangzhou-Shenzhen-Hong Kong-Guangdong-Hong Kong-Macao Greater Bay Area information linkage absolute control advantage, four cities Foshan, Dongguan, Zhuhai, Macao regional hub position steadily highlighted. The entire information connection network of the urban agglomerations tends to be flat and polycentric at the same time. (3) The regional core-edge hierarchy is well established, with the four cities of Guangzhou, Dongguan, Shenzhen, and Hong Kong creating a northwest-southeast orientation. The core metropolis regions of Guangdong, Hong Kong, and Macao in the Greater Bay Area increasingly exert a radiation spreading effect to the northeast and southwest. (4) The urban economy, transportation distance, and information infrastructure have substantial effects on the information connection intensity network of urban clusters.

A hybrid forecasting architecture for air passenger demand considering search engine data and spatial effect

As the global process of digitalization accelerates and breakthroughs in internet technology emerge, governments worldwide are advocating for data-driven decision-making, aiming to enhance public service efficiency and stimulate economic growth. Against this backdrop, this study focuses on utilizing search engine data (SED) to improve air passenger demand forecasting, responding to national policies aimed at enhancing data analysis capabilities and promoting the development of intelligent transportation systems; however, the existing research is confined to the exploration of the temporal dependency between SED and air passenger demand variables with ignoring the spatial dependency. In order to eliminate this blind spot and catch from various parts of tourist attention, this study proposes a novel SED-driven hybrid forecasting architecture inspired by the theory of spatial effect between adjacent tourist destinations. The architecture includes three main steps: (1) construction of spatial-temporal SED variables, based on two-stage data preprocessing method; (2) variable decomposition and reconstruction, based on TVF-EMD algorithm; (3) prediction of different components of air passenger demand, employed ARIMA model and IHGS-KELM based multi-model fusion strategy respectively, where the IHGS algorithm integrates the circle chaos initialization strategy and the nonlinear convergence factor strategy. To confirm the practical applicability of this hybrid architecture, five comparative experiments based on the actual dataset are designed. The principal results are concluded as follows: (1) spatial-temporal SED is conducive to a fairly accurate air passenger demand forecasting; (2) the multi-model fusion strategy can integrate the fortes of various types of prediction models to obtain better prediction accuracy; (3) the adaptive ensemble method based on IHGS-KELM can contribute to the upgradation of prediction performance of air passenger demand.

P045 Google Internet Searches Related to Inflammatory Arthritis: an Observational Study Using Google Trends Data

Abstract Background/Aims Inflammatory arthritis (IA) care involves a paradox. Rheumatologists focus on reducing disease activity through treat-to-target (T2T). But surveys/qualitative studies show patients’ main concern is pain. As these studies include a minority of patients, their findings may not be generalisable. The internet has transformed how patients access health information. Examining search engine data provides an alternative method to understand the health areas of most importance to people with IA. We examined publicly available Google search engine data, to determine the relative popularity of searches for different aspects of health made in conjunction with IA. Methods We used Google Trends data from 2011-2022 to determine the relative popularity of searches for “patient symptom” (pain; work; stiffness; fatigue; mood) and “T2T” (DMARDs; inflammation; synovitis; corticosteroids) domains made with rheumatoid arthritis (RA), psoriatic arthritis (PsA), and ankylosing spondylitis (AS) in the UK/USA. Google Trends normalises a sample of search data by popularity over time and region, presenting it on a 0-100 scale at each time-point (100 represents the time-point with most searches). It allows comparisons of up to five search term combinations. We evaluated >100 search term combinations to ascertain the most popular “patient symptom” and “T2T” domains (comparing mean normalised scores across time-points). Results Patient Symptom Domains: in all IA forms in the UK/USA, pain was the most popular term. Its popularity increased over time. UK searches for pain gave mean scores of 58, 34, and 39 with RA, PsA, and AS. UK searches for other patient symptom domains gave mean scores ranging 2-7 with RA; no scores were generated with PsA/AS. In the USA, pain gave mean scores of 79, 51, and 63 with RA, PsA, and AS, compared to scores ranging 2-6 for other domains. T2T Domains: methotrexate was the most popular search term with RA and PsA in the UK (mean 28 and 21) and USA (mean 63 and 33); it was only slightly more popular than searches for synovitis and inflammation. For AS, in the UK and USA inflammation was most popular (mean 9 and 34). Most Popular Patient Symptom vs Most Popular T2T Domain: for RA, PsA, and AS, pain was substantially more popular in the UK (mean 58, 30, 52) and USA (mean 79, 54, 63) than methotrexate in RA and PsA in the UK (mean 16 and 14) and USA (mean 15 and 11) and inflammation in AS in the UK (mean 6) and USA (mean 8). Conclusion Pain is the most popular search term used with IA in Google searches in the UK/USA, supporting surveys/qualitative studies showing pain is the health aspect of most importance to people with IA. Routine pain assessments should be embedded within T2T strategies to ensure patient perspectives are considered. Disclosure M. Akthar: Grants/research support; ‘M.A. previously received funding by the National Institute for Health and Care Research (NIHR Academic Clinical Fellowship in Rheumatology)., The views expressed are those of the author and not necessarily those of the NIHR, NHS or the Department of Health and Social Care.’. K.J. Mason: Grants/research support; ‘K.J.M. is funded by the National Institute for Health and Care Research (NIHR; Grant Reference Number NIHR203281)., The views expressed are those of the author and not necessarily those of the NIHR, NHS, or the Department of Health and Social Care.’. I.C. Scott: Grants/research support; ‘I.C.S. is funded by the National Institute for Health and Care Research (NIHR Advanced Research Fellowship, NIHR300826)., The views expressed are those of the author and not necessarily those of the NIHR, NHS, or the Department of Health and Social Care.’

Semiconductor Price Index Predicting Based on a Novel Improved AdaBoost Feature-Weighted Combination Model

The semiconductor price index serves as a vital metric for assessing technological developments and related market trends. Establishing a more accurate forecasting model for the semiconductor price index is of significant importance for analyzing the industry’s trends and market directions. In this paper, a novel framework for semiconductor price index forecasting is proposed. In addition to traditional financial data, the study introduces search engine data (Google Trends) representing investor attention, and introduces text information extracted from online news headlines reflecting major market events and government policies as independent variables. Used to predict the dependent variable: The PHLX Semiconductor Sector (SOX). First, the XGBoost model is employed to compute the importance scores of each feature. Then, a feature weight coefficient indicator is constructed based on these importance scores to calculate the weight coefficient indicator values for each feature. These indicator values are then used to weight the kernel function of Support Vector Regression (SVR), resulting in weighted Support Vector Regression (WSVR). Finally, WSVR is utilized as the base learner for Adaptive Boosting (AdaBoost), yielding the XGBoost–WSVR–AdaBoost model based on feature weighting. The proposed model outperforms AdaBoost, RNN, ERT, LSTM, and other models in terms of Mean Absolute Percentage Error (MAPE) and goodness-of-fit (R2\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${R}^{2}$$\\end{document}). It also exhibits superior predictive performance compared to models in ablation experiments, and the introduction of text data or Google trends further improves the prediction performance of the model. In conclusion, the improved AdaBoost feature-weighted combination model proposed in this study offers a more accurate prediction for semiconductor price index.

Forecasting tourism demand with search engine data: A hybrid CNN-BiLSTM model based on Boruta feature selection

Using search engine data (SED) to forecast tourist flow is essential for management and security warnings at tourist attractions. Existing prediction models cannot effectively handle noise in the SED and external uncertainties. Thus, insufficient feature extraction may hinder the fitting of tourist flow time series. To improve the prediction accuracy, a forecasting method combining the Boruta algorithm (Boruta), bidirectional long short-term memory (BiLSTM), and a convolutional neural network (CNN) is proposed in this study. The model was tested using a monthly tourist flow dataset of three adjacent tourist attractions (Hongcun Village, Mount Huangshan, and Xidi Village) in Anhui Province, China, from January 2011 to March 2021. Forecasting for Hongcun Village shows that the proposed model performs best in three-step-ahead forecasting with a root mean square error (RMSE) of 2341.77 and a mean absolute percentage error (MAPE) of 2.85 %. A comparative analysis demonstrated that the hybrid CNN-BiLSTM model outperformed benchmark models. Meanwhile, Boruta reduced the RMSE and MAPE of the model by 59.93 % and 54.11 %, respectively, which were superior to those of the baseline methods. Furthermore, a robustness test was applied considering Mount Huangshan and Xidi Village as additional exercises for prediction. The robustness test confirmed that the proposed model has the proper generalization capability. Accordingly, the forecasting framework used in this study is robust and operational. This can facilitate a new approach for tourism demand forecasting with SED and assist operators in making managerial decisions.

Development of a novel dynamic nosocomial infection risk management method for COVID-19 in outpatient settings

BackgroundApplication of accumulated experience and management measures in the prevention and control of coronavirus disease 2019 (COVID-19) has generally depended on the subjective judgment of epidemic intensity, with the quality of prevention and control management being uneven. The present study was designed to develop a novel risk management system for COVID-19 infection in outpatients, with the ability to provide accurate and hierarchical control based on estimated risk of infection.MethodsInfection risk was estimated using an auto regressive integrated moving average model (ARIMA). Weekly surveillance data on influenza-like-illness (ILI) among outpatients at Xuanwu Hospital Capital Medical University and Baidu search data downloaded from the Baidu Index in 2021 and 22 were used to fit the ARIMA model. The ability of this model to estimate infection risk was evaluated by determining the mean absolute percentage error (MAPE), with a Delphi process used to build consensus on hierarchical infection control measures. COVID-19 control measures were selected by reviewing published regulations, papers and guidelines. Recommendations for surface sterilization and personal protection were determined for low and high risk periods, with these recommendations implemented based on predicted results.ResultsThe ARIMA model produced exact estimates for both the ILI and search engine data. The MAPEs of 20-week rolling forecasts for these datasets were 13.65% and 8.04%, respectively. Based on these two risk levels, the hierarchical infection prevention methods provided guidelines for personal protection and disinfection. Criteria were also established for upgrading or downgrading infection prevention strategies based on ARIMA results.ConclusionThese innovative methods, along with the ARIMA model, showed efficient infection protection for healthcare workers in close contact with COVID-19 infected patients, saving nearly 41% of the cost of maintaining high-level infection prevention measures and enhancing control of respiratory infections.

The impact of search engine data sharing on competition and consumer welfare

ABSTRACT Search engines match user queries with webpages. Search efficiency, increases when data-driven network effects attract more users and generate economies of scale and scope in data aggregation. Network effects enhance user welfare but also trigger antitrust concerns because they may “tip” the market towards a single dominant search engine – Google Search. The EU Digital Markets Act imposes an obligation on very large gatekeepers search engines to share their user data with smaller search engines to facilitate market entry and competition. The available empirical literature on search engine efficiency shows that asymmetric data sharing, from larger to smaller search engines, may increase competition but reduce user welfare because it fragments the user data pool that is crucial for efficiency, in particular for rare queries. Tension between competition and user welfare can be overcome by symmetric and mutual data sharing between all search engines, irrespective of market share.

Cryptocurrency interest in geographical regions with high unemployment rates

Purpose- In regions with high unemployment rates, individuals often face economic uncertainty and may seek alternatives to traditional financial systems. In such circumstances, some individuals may turn to alternative investments and financial instruments, with cryptocurrencies such as Bitcoin being among them. The value of these cryptocurrencies is often volatile due to their speculative nature, which can lead investors to take on more significant risks. While this situation can result in substantial gains, it can also lead to significant financial losses. The primary use of cryptocurrencies is generally for investment purposes. Since cryptocurrencies are digital, they operate entirely in a virtual environment. The loss or theft of the digital password in the virtual space means the loss of the wallet. Security in cryptocurrencies is very weak. The lack of government guarantees and legal infrastructure globally poses certain threats to investors, making cryptocurrencies carry a higher risk compared to other investment instruments. This study will investigate the regional relationship between cryptocurrency interest and unemployment rates. The aim is to determine whether individuals living in geographical areas with high unemployment rates show more interest in cryptocurrencies compared to those in other geographical regions. Methodology- In line with the research objective, Google search engine data for the past 10 years has been analyzed using content analysis methodology. A query has been conducted to determine in which regions of Turkey cryptocurrency names are most frequently searched, and a comparison has been made on a provincial level across all 81 provinces. Additionally, considering the unemployment data provided by the Turkish Statistical Institute (TUIK), regions with high and low unemployment rates have been identified. Subsequently, an investigation has been carried out to determine whether there is any correlation between regional Google searches and unemployment rates in these areas. Findings- When examining the findings from the perspective of unemployment, it has been observed that the regions where cryptocurrency names are most frequently searched align with areas having the highest unemployment rates, while the least searched areas correspond to regions with lower unemployment rates. Among the 10 provinces with the highest unemployment rates, 8 of them are included in the list of 10 provinces with the highest Bitcoin search frequency. Similarly, 8 out of 10 provinces with the lowest employment rates are also listed among the 10 provinces with the highest Bitcoin search frequency. Furthermore, among the 10 provinces with the lowest per capita gross domestic product (GDP), 7 of them are included in the list of 10 provinces with the highest Bitcoin search frequency. Conclusion- As a result of the relationship analyses conducted in the research, it has been observed that in Turkey, regions with high unemployment rates, low employment rates, and low Gross Domestic Product (GDP) have a higher frequency of Bitcoin searches compared to other regions. Keywords: Cryptocurrency, unemployment, bitcoin JEL Codes: P44, E24, E44

The impact of search engine data sharing on competition and consumer welfare

The impact of search engine data sharing on competition and consumer welfare

Predicting Tourism Demand by Combining Search Engine Data

Predicting Tourism Demand by Combining Search Engine Data

The Drug War and Regional Social Capital in Mexico

AbstractThe onset of the war on drugs in Mexico at the beginning of 21st century had far‐reaching effects on its citizens, including most obviously, an unprecedented increase in the homicide rate. We analyse the correlation between violence on social capital in the 32 federal states of Mexico from January 2004 to December 2016. Given the lack of data in the conflict regions of Mexico, we apply the indirect approach proposed by Guriev and Melnikov (2016), which uses internet search engine data to proxy social capital. Our results show a negative relationship between violence and social capital in Mexico. Moreover, we document a positive spatial correlation for social capital. Overall, we present an example of how the analysis of internet‐based data can contribute to the understanding of socioeconomic developments in conflict regions with unreliable standard data.

Social Trend Mining: Lead or Lag

This research underscores the profound implications of Social Intelligence Mining, notably employing open access data and Google Search engine data for trend discernment. Utilizing advanced analytical methodologies, including wavelet coherence analysis and phase difference, hidden relationships and patterns within social data were revealed. These techniques furnish an enriched comprehension of social phenomena dynamics, bolstering decision-making processes. The study’s versatility extends across myriad domains, offering insights into public sentiment and the foresight for strategic approaches. The findings suggest immense potential in Social Intelligence Mining to influence strategies, foster innovation, and add value across diverse sectors.

Investor sentiment based on search engine data for predicting stock returns in Indonesia industrial sector

This study aims to investigate investor sentiment and its effect on the industrial sector in Indonesia. In the study, investor sentiment was extracted from data obtained from search engines. Then the data were used to see how the sentiment affected the stock return in each listed industrial company. Fifteen industrial sector companies listed on the stock exchange were selected; each was analyzed for their search volume and the effect on stock returns using panel data regression from March 2020 to April 2022. The result shows that investor sentiment affects the level of stock returns in the industrial sector in Indonesia. This result indicates that with the increase in search volume on search engines, which reflects positive sentiment, there will be an increase in stock trading transactions and vice versa. This study's findings will help investors make investment decisions, especially in the industrial sector.

Complementarities in learning from data: Insights from general search

The ability to make accurate predictions relating to consumer preferences is a key factor of a digital firm's success. Examples include targeted advertisements and, more broadly, business models relying on capturing consumers' attention. The prediction technologies used to learn consumer preferences rely on consumer generated data. Despite the importance of data-driven technologies, there is a lack of knowledge about the precise role that data-scale plays for prediction accuracy. From a policy perspective, a better understanding about the role of data is needed to assess the risks that “big data” might pose for competition. This article highlights potential complementarities between different data dimensions in algorithmic learning. We analyze our hypothesis using search engine data from Yahoo! and provide evidence that more data in the within-user dimension enhances the efficiency of algorithmic learning in the across-user dimension. Our findings suggest that ignoring these complementarities might lead to underestimating scale advantages from data.