Measuring economic activity in China with mobile big data

We assume that the numerical changes of employees and consumers could be tracked from geo-positioning data, by which we construct indices to measure economic activity. In order to verify this idea, we study four cases with obvious employment volatility: two announced mass layoffs, and the remaining two experienced rapid developments with an employment surge. We first identify employees and count the monthly number in these areas by method mentioned in Section 3.2.

Figures 2A-C show the first two cases. Firm I is a large shoe factory located in a south-eastern coastal city in Guangdong Province, the heart of China’s export manufacturing. Because of global industrial transfers, increasing labour costs and reduced export orders, the plant was closed in the first quarter of 2016, leaving thousands of workers jobless. As shown in Figure 2C, there was a sharp drop in the number of employees (the dark blue line) during this period (the grey box). A similar shock hit Firm II, a mobile phone factory located in Suzhou, Jiangsu Province (Figure 2B), when it was shut down at the end of December 2015, with hundreds of workers losing their jobs. It is worth noting that Firm I gradually laid off its employees over three years, whereas Firm II shut down suddenly at the end of 2015. The discrepancy between these two cases is also mirrored in our results, as shown in Figure 2C: Employment for Firm I declined 20% further in 2015 from 2014, with a consequent year-over-year decline of 30% in the first quarter of 2016. By contrast, employment for Firm II remained stable and then dipped dramatically in the last quarter of 2015, after the shutdown.

Figures 2D-F show the second two cases of increasing employment. Zone III is a software park in Beijing (Figure 2D), home to many high-tech companies. Most of these companies began to relocate there in 2014, and since then, as shown in Figure 2F (red line), the number of employees has almost doubled. Firm IV is a fast-growing start-up based in Beijing (Figure 2E) that has experienced rapid expansion since 2015. As the figure shows, its employment jumped in the second quarter of 2015 after it raised billions of dollars in investment funding.

We build an employment index to track macro trends in employment by aggregating employment changes in 2,000 labelled AOI, including 1,000 manufacturing industrial parks and 1,000 high-tech industrial parks. Since we sample users with consistent location points during a 13-month rolling window, the first 13-months’ index is simply the normalized number of workers in AOIs. Starting from the 14th month, the employment index is calculated by the following equations: where \(\mathit{rate}_{t}\) is the year-over-year percentage change of number of workers within all AOIs, and \(\mathit{employment}_{t}\) is the employment index of month t.

Figure 3A illustrates the monthly employment index from January 2014 to June 2016 without seasonal adjustment, in which the grey, blue and red lines indicate the employment index for all industrial parks, manufacturing parks, and high-tech parks, respectively. All indices are normalized by the mean value for the year 2014 (set to be 100). Figure 3B shows the corresponding year-over-year growth rate. The following findings can be drawn from the figures: (1) The overall employment index of industrial parks was relatively steady, with a slow decline over the last two years; employment began falling in the beginning of 2016, indicating a general slowdown in the economy. (2) The monthly employment index of manufacturing parks showed a sustained and more obvious decline in overall performance; this downward trend is similar to what was found in the official manufacturing purchasing managers index (see Supplementary Figure S4). The employment index of manufacturing parks dipped more than 5% year-over-year during most of the study period, indicating that this sector seriously suffered from the economic slowdown. (3) The employment index of high-tech parks, by contrast, saw a healthy and continuous increase, with more than a 3.5% year-over-year increase over the last two years, which suggests that high-tech companies in industries such as software and bio-tech (referred to as belonging to the ‘new economy’) continued to create more jobs than other sectors.

In a similar fashion, we track changes in consumer activity for approximately 4,000 major commercial areas, covering most of the large shopping malls in China. We then construct a monthly consumer index to measure the changes and track the growth in consumer activity in these areas using Eq. 1 and Eq. 2. The consumer index generally climbs during the summer holidays (July and August) and before the Spring Festival in December and January and then declines during the Spring Festival in February (Figure 4A). One can see that the consumer index gradually inches down and that its year-over-year growth rate fluctuates, with more dips below zero than spikes over time, indicating weakening consumer demand (possibly dragged down by the overall slowdown and by e-commerce). The year-over-year growth rate shown by the consumer index (Figure 4B) began rising in September 2015 and peaked at approximately 9% in December; however, at the beginning of 2016, it fell back below zero and sank to nearly −7% in June of this year, indicating relatively low consumption demand in the first half of 2016.

To investigate whether consumer index matters for economic performance, we evaluate our index by comparison with firm level revenue. We collect monthly revenue data of three large commercial areas, located in Beijing, Shanghai, and Shenzhen respectively. For comparison, we rescale the consumer index of each commercial area, and plot them together with corresponding revenue data. Figure 5 shows that our consumer index is highly correlated with the revenue data, and the Pearson correlation coefficient is 0.94 for each city, suggesting that our consumer index could be a reasonable proxy for measures of commercial performance.

In summary, the proposed employment index and consumer index measure the economic performance of selected industrial parks and commercial areas in China, depict the complicated reality of China’s economy, and provide a new perspective for charting macro-economic performance from a bottom-up view.

Foot traffic is one of the most critical metrics for service sector businesses such as retail stores, restaurants, movie theatres, and hotels. We observe that, unsurprisingly, the volume of map queries regarding one specific location is highly correlated with that of offline foot traffic. This correlation is illustrated in Figure 6, which shows the number of people searching for Shopping Malls I and II (orange dashed line) on Baidu Maps and the foot traffic (red solid line) of each place (estimated from the geo-positioning dataset). In both cases, map queries and foot traffic present obvious periodical patterns, peaking on weekends and dipping during weekdays. To evaluate the efficiency and accuracy of the foot traffic estimation via map query data, we compare two seasonal autoregression (AR) models to predict foot traffic: where \(y_{t}\), \(y_{t-1}\), and \(y_{t-7}\) are the numbers of consumers at time t, \({t-1}\), and \({t-7}\), respectively. Here we add \(y_{t-7}\) to the regression because human activities have weekly periodicity. The volume of map queries at time t is \(q_{t}\), and \(e_{t}\) is the error term. The regression results, shown in Table 1, indicate that adding map queries to the regression significantly improves the in-sample fit. The \(R^{2}\)s of Shopping Malls I and II increase from 0.727 and 0.767 to 0.914 and 0.931, respectively. In fact, if we use only map queries to fit the foot traffic estimation, the results are surprisingly accurate: the \(R^{2} \)s are 0.880 for Shopping Mall I and 0.910 for Shopping Mall II (Figure 6CF). Therefore, we can use map queries to estimate the volume of foot traffic and evaluate the performance of foot-traffic-dominated stores, firms, and industries in the service sector.

Here, we demonstrate two cases to validate the potential power of map queries for estimating and forecasting revenue. The first one is a revenue prediction for Apple Inc. (Apple) in China, and the second is box office ‘nowcasting’ and fraud detection.

As a listed company, Apple reports its earnings results to the public quarterly with approximately one month’s lag time. According to official numbers, Apple reported $75.9b in revenue during the last quarter of 2015 (note that the fiscal quarter is different from the calendar quarter), and China, the second-largest market globally, contributed $18.4b (in the same quarter a year prior, the total revenue was $74.6b, and the Chinese market contributed $16.1b). In the first quarter of 2016, Apple’s revenue ($50.6b) was down year over year for the first time since 2003, and its revenue in China declined by approximately 26%.

Given that we are able to estimate foot traffic volumes for offline stores by using map queries, should we conclude that these volumes are related to reported revenue? Is it possible for us to predict quarterly revenue prior to earnings announcements?

To answer these questions, we first select a list of flagship Apple Stores in Mainland China (see Additional file 1 for the list) and then count the volume of map queries for all the stores. As shown in Figure 7A, the blue line spans from the last quarter of 2014 to the first quarter of 2015; the red line spans the corresponding time period for 2016. We also calculate the year-over-year changes in map queries (the grey area in Figure 7A) and compare it with revenues (Figure 7B). As Figure 7 shows, the volume of map queries showed a 15.4% year-over-year increase in the last quarter of 2015 and a 24.5% decline in the first quarter of 2016, and revenues similarly showed a 14% increase and a 26% decrease during the corresponding period. The impressively strong correlation indicates that map query data allow us to ‘nowcast’ the company’s revenues and reveal future trends. We then applied this method to map query data of the second quarter of 2016 and projected that Apple’s revenue in Greater China in this quarter may decline 34% on a year-over-year basis (one month prior to earnings announcements). The final number in the earnings reports was 33%, which is close to our prediction.

Going beyond estimation at the firm level, we find that map query data are also predictive at the level of the specific consumer sector. To better illustrate, we collect daily box office data during 2015 from the China Box Office Database [58921.​com] and then build models to ‘nowcast’ daily box office revenues by using map queries related to cinemas. Here, we use the word ‘nowcasting’ instead of ‘forecasting’ because we use map query data to predict box-office revenues for the same day, which is approximately one day earlier than the official statistics. To reduce the effect of user search trends, we normalize the volume of daily map query data and use it as the independent variable in our regression models. Our baseline model is a simple AR model with 1-day and 7-day time lags, as we have no further information about movies (e.g., production budgets and number of opened screens, which are used as input features in ref. [12]). The regression equations are similar to Eq. 3 and Eq. 4.

As shown in Supplementary Table 1, map queries for cinemas significantly improve the in-sample fit: the \(R^{2}\) of the baseline model is 0.489, while that of the map query model is 0.934. To further investigate the models’ out-of-sample performance, we apply a rolling window forecast similar to the method used in [10, 49]. Given the selected date (July 1st, 2015, in this case), prediction in the \((t-1)\) time step is based on estimates of all previous time periods (≤t).

The results are plotted in Figure 8, where the grey line represents the actual box office statistics, the blue line represents the forecasting results of the baseline model, and the red line denotes the forecasting results of our map query model. The absolute error of the baseline model is almost always larger than that of the map query model. The mean absolute error (MAE) of the map query model declines 67% below that of the baseline model, with a value of 13.2 million RMB (the mean of daily box office is 118 million, with a standard deviation of 75.8 million), as shown in Figure 8B.

Further, we reveal some additional insights when analysing data for specific movies and find that map queries are helpful for detecting box office fraud. For example, Monster Hunter, a Chinese box office champ in 2015 earning approximately 2.4 billion RMB in revenue, was reported to have artificially inflated its box office by scheduling midnight screenings with almost no audience present in cinemas owned by the movie’s producer [50]. To check for such inflated numbers, we select a list of suspected cinemas owned by the producer of Monster Hunter (approximately 30 cinemas) and a list of 30 other cinemas as a control group. We then calculate the volume of map queries for these locations from July 2015 to September 2015, when the movie was shown.

Figure 9A shows the results of suspected cinemas, and Figure 9B shows those of the control group. The grey lines and red lines indicate the total box office revenues and Monster Hunter box office revenues, respectively. The blue lines show the fitting results by map queries, and the black dashed lines denote forecasting errors. Comparing the forecasting errors for two sets of selected cinemas, one can see that the forecasting error for the control group is small and stable, whereas for the suspected cinema set, there is a sharp increase in the forecasting error from August 25 to September 16 (marked by the grey box). Such an anomaly indicates a high probability of box office fraud.

As demonstrated above, the volume of map queries is strongly correlated with foot traffic and is able to estimate offline consumer spending. We therefore construct the consumption trends for various industries by aggregating the map queries belonging to the same POI category. For example, one who searches for a car dealership on a map will be regarded as a potential automobile industry consumer. In the main text, we present consumption trends for four sectors, Automobile Sales, Restaurants, Finance & Investments, and Tourism, and compare our results with several other indicators (see Supplementary Figures S5-S7 for more sectors’ trends). All trends are normalized by the mean value for the year 2014.

Figure 10A shows the overall rising trend in automobile sales based on an analysis of approximately 190,000 related locations, which is highly correlated with the sales data published by the Automobile Association of China [http://​www.​caam.​org.​cn/​]. The Pearson correlation is 0.913, with \(p < 0.001\), indicating that the proposed trend can effectively track market changes in real time.

Figure 10B shows the relative declining trend in the restaurant sector, which includes almost 3 million restaurants. We compare our result with the Tsinghua UnionPay Advisors’ Restaurant & Catering (TUPARC) Index [31], a consumer spending indicator based on bank card transaction data from the China Cuisine Association’s 100 top-ranked restaurants. While our proposed trend is consistent with the TUPARC over most of the time period (the Pearson correlation coefficient is 0.764, with \(p < 0.001\)), our restaurant consumption trend began dipping in 2016 and fell to a new low of 85 in June.

Figure 10C shows results for the finance and investment sector, which includes 230,000 related locations. Interestingly, we find that our proposed trend resembles the trends of the Shanghai Composite Index, one of the stock market indices in China. Our on-going investigation indicates that the proposed trend may be a leading indicator for stock market movements, especially for ‘turning points’ (marked with circles). Specifically, we follow the trading strategy proposed in [13] to analysis weekly changes in map query volumes for investment sector (e.g. banks, security companies). If the query volume of week t (note as \(q_{t}\)) overs the average volume of three weeks before, we would take the ‘long position’: buy the stock index at the closing price on the first trading day of week \({t+1}\) and sell at the closing price on the last trading day of week \({t+1}\). Otherwise, we would take the ‘short position’: sell at the closing price of the first trading day and buy at the last trading day of week \({t+1}\): We apply this weekly strategy to the stock market dataset in the period between Jan. 2014-Apr. 2016 (120 weeks). If we buy at the beginning week (the Shanghai Composite Index is 2038) and hold until the end (the Index is 2956), the yield is approximate 17.5% annually. In contrast, if we take the ‘long/short strategy’ as mentioned above, the yield is approximate 49.3% annually (transaction fees are neglected in this analysis).

Figure 10D shows the trend for the travel sector, which comprises 300,000 tourist attractions. The overall trend grows moderately, exhibiting obvious spikes during major Chinese holidays, such as the Spring Festival, the National Holiday and the summer holiday, which generally lasts from July to August.

Our consumption trends for these sectors demonstrate that, on a year-over-year basis, automobile sales rose, restaurant spending declined, finance investment was related to stock market volatility, and travel spending grew moderately. All these findings reflect the complicated reality of China’s service economy.