Analysis of the tax sensitivity of particular sectors of the economy

Gerasimova Anna Evgen'evna ORCID: 0000-0001-8480-6279 PhD in Economics Associate Professor; Department of Taxes and Tax Administration; Financial University under the Government of the Russian Federation 15 Verkhnyaya Maslovka str., Moscow, 127083, Russia kharitonova.ae@yandex.ru Other publications by this author

Introduction

Sector tax sensitivity analysis is the process of assessing the impact of changes in tax policy on economic activity, profitability and resource allocation in various sectors of the economy. It reflects changes in tax rates or tax structures that may affect the productivity and competitiveness of individual industries, as well as overall macroeconomic stability. In the process of development, there are more and more differences between industries every year in terms of the technological level of production development, workers' qualifications, etc. As a result, there is a tendency to increase the differentiation of tax regulation instruments by industry.

The sectors of the economy differ significantly from each other in terms of tax sensitivity. It is believed that "the deeper the level of processing and, accordingly, the higher the level of manufacturability and complexity of production, the more sensitive the industry is to tax exemptions" ^[1].

The aim of the study is to assess the tax sensitivity of individual sectors of the economy using modern econometric methods and machine learning methods for making managerial decisions differentially for each individual sector.

The main stages of the study of the tax sensitivity of economic sectors include:

- to analyze the tax burden by industry;

- analyze the structure of payment receipts to the consolidated budget by industry;

- to build and compare models for forecasting the tax burden for selected organizations in various sectors of the economy;

- evaluate the results obtained.

The scientific novelty of the research lies in the development of an author's methodological approach based on artificial intelligence methods to assess differences in the tax sensitivity of individual sectors of the economy for making managerial decisions differentially for each individual sector.

Literary review

According to the "Forecast of socio-economic development of the Russian Federation for 2024 and for the planning period of 2025 and 2026" in 2023, the global economy continues to slow down, but the prospects for the development of the Russian economy are more optimistic. The forecast notes that "by the end of 2023, economic growth will not only compensate for the recession of last year, but also achieve an increase to the level of two years ago" ^[2]. At the same time, the most promising sectors of the real sector, especially the manufacturing industry, are singled out.

In order to achieve stable economic growth, a balanced fiscal policy is needed that would support the economy with sufficient stability. To ensure a relatively high level of government spending to support economic development, the tax burden may be unevenly distributed across industries. At the same time, taxes can have both a positive impact on economic development due to increased government support, and reduce growth rates due to an increase in the tax burden ^[2].

The tax burden has significant differences both by industry and by region of the country. Thus, in ^[3], regional peculiarities were identified depending on their resource potential. At the same time, it is noted that the financial situation of organizations in different industries does not always depend on the tax burden. Industries with a low tax burden often have a worse financial condition than industries with a high tax burden. But regional differences depend primarily on the development of individual industries, which are conditioned by the prevailing conditions.

There are several approaches to the construction of models for determining the tax sensitivity of individual industries: V.G. Papava ^[4], G.G. Loladze ^[5], E.V. Balatsky ^{[6, 7]}, Yu.Sh. Ananiashvili ^[8], etc. However, due to the significant changes that have occurred in the economy during the spread of coronavirus infection and under international sanctions, dynamic models are becoming inapplicable. Time series have significant structural shifts, factors become unreliable and as a result it becomes impossible to assess tax sensitivity.

To assess the tax burden, a number of authors use correlation and regression analysis ^{[9, 10, 11]}, which allows building multifactorial models to assess the impact of individual factors and forecasting. The advantages of this method are the use of static series, the ability to isolate the independent influence of factors and the high quality of forecasts. However, in order to build a model, a large number of conditions must be fulfilled (the absence of multicollinearity of factors, a sufficient number of initial values, the fulfillment of the prerequisites of the least squares method, etc.). As a result, it is quite difficult to choose a linear combination of factors that would satisfy all conditions and at the same time the quality of the models would be quite high. The nonlinear nature of the relationships between economic variables is also a problem, which makes it difficult to select a regression model. With a large set of initial information, machine learning models have advantages, which can build nonlinear models and identify unusual relationships between variables. Despite the popularity of classical econometric methods, the construction of dependence models must be compared with modern machine learning methods to obtain reliable and reasonable results.

As a result, it becomes necessary to search for new ways to identify intersectoral differences in the size of the tax burden, taking into account the modern development of artificial intelligence systems and work with big data.

Materials and methods of research

The analysis of the sensitivity of economic sectors to the tax burden can be carried out in several stages (Figure 1).

Figure 1. – Stages of the analysis of the sensitivity of economic sectors to the tax burden

Source: compiled by the author

At the first stage of the work, data is collected at the macroeconomic level in dynamics, characterizing the level of tax burden and the structure of revenues to the consolidated budget of the Russian Federation according to the Federal State Statistics Service and the Federal Tax Service.

At the second stage of the analysis, the assessment of the sectors of the economy with the highest volumes of tax revenues to the consolidated budget of the Russian Federation is carried out. For the analysis, it is proposed to consider the industries with the highest tax revenues, since their role in budget formation is the highest. After selecting these industries, it is necessary to return to stage 1 and select data on organizations in the relevant industries to assess factors and build forecasting models. The data for analysis is unloaded from the SPARK system ^[12] and their primary processing is carried out: missing values, outliers are deleted, and the indicator system is calculated. At the 2nd stage, according to these indicators, the variation of the feature, the relationship between the variables and the possibility of using it for forecasting purposes are evaluated.

At the 3rd stage of the analysis, models for forecasting the tax burden are built using econometrics and machine learning methods.

The methods of constructing models were used:

- The Decision Tree method, which consists in dividing the initial population into relatively homogeneous groups and allows you to find atypical patterns in the data and make forecasts of sufficiently high quality. The division of the population is carried out in several stages (depending on the depth of the tree) and according to the most significant factors. The verification condition is designated as a node of the tree, and the branches of the tree are those units that fulfill the specified conditions. As a result of splitting the tree, they come to the end nodes – this is the result of forecasting ^[13].

- Gradient boosting (AdaBoost) is a machine learning method based on the construction of decision trees, while at each stage the model is improved by compensating for the shortcomings of the previous ones based on the optimization of an arbitrary differentiable loss function ^[14]. This method also allows you to build non-trivial relationships between variables and, in terms of model quality, often surpasses many other machine learning methods.

- The Nearest neighbor method is based on the idea of the proximity of objects in a multidimensional feature space ^[15]. The essence of the method is that the value of a new effective feature is determined by the values of the factors closest to it. The predicted values are determined by the closest data sections according to a certain proximity function ^[16]. The model preserves the trends and patterns of the original population and applies them to new values to search for forecasts.

Linear regression is an econometric analysis method that allows you to identify linear relationships between variables, and is also used as a machine learning method. The algorithm builds linear models between 1 dependent variable y and several independent variables, considered as factors ^[17]. The model gives high results if there is a linear relationship between the variables.

Based on the results of building forecasting models, it is possible to draw conclusions about the tax sensitivity of individual industries.

The results of the study

The tax burden by economic sector is calculated as the ratio of revenues to the consolidated budget of Russia to the gross value added of the relevant industry (Table 1).

Table 1 – Tax burden by economic sector, 2018-2023, %

Source: compiled by the author according to the data of the Federal State Statistics Service and the Federal Tax Service

According to the table, it can be seen that the structure of revenues to the consolidated budget of the Russian Federation by sectors of the economy does not change significantly. The linear coefficient of structural shifts over the years was respectively: 0,9; 1,1; 1,3; 1,9; 2,2. Thus, it can be noted that the largest differences in receipts occurred from 2022 to 2023.

By industry, the highest tax burden is observed in the following industries: mining (annual average 61.9%), information and communication activities (28.1%), manufacturing (27.3%) and provision of electric energy, gas and steam; air conditioning (27.3%).

The lowest tax burden by industry is typical for the following industries: agriculture, forestry, hunting, fishing, fish farming (3.8%), public administration and military security; social security (6.4%), real estate transactions (6.9%) and activities in the field of health and social services (9.4%).

For the Russian economy, the main role is played by the volume of revenues to the consolidated budget from various industries (Table 2).

Table 2 – Structure of payment receipts to the consolidated budget of the Russian Federation by economic sector, 2018-2023, %

Source: compiled by the author according to the Federal Tax Service

In the structure of receipts of payments to the consolidated budget of the Russian Federation by branches of the economy, the linear coefficient of structural shifts (from 0.3 to 0.99 by year) indicates the uniformity of receipts by year.

The leaders in the structure of budget revenues are: mining (34.2%), manufacturing (17.2%), wholesale and retail trade; repair of motor vehicles and motorcycles (11.7%), financial and insurance activities (5.1%) and construction (4.2%). It is for these industries that it is most important to assess the sensitivity to the tax burden.

At the 3rd stage of the analysis, a sample of organizations of the relevant sectors of the economy from the Spark system for 2023 was made ^[12]. Data processing was performed using the python programming language with the Anaconda distribution in the Jupyter Lab environment and its libraries: pandas, numpy, seaborn, matplotlib and. The uploaded data was cleared of missing values (columns with more than 90% of omissions were deleted. And then the organizations for which at least 1 value was omitted were deleted).

To build models, a system of indicators characterizing the economic and financial situation of organizations was calculated:

- Tax burden, %;

- Age of the company, years;

- The share of non-current assets in total;

- Revenue per 1 ruble of assets;

- The share of non-current assets in total;

- The ratio of accounts receivable to the company's assets, %;

- The share of working capital in the company's assets, %;

- Net debt to equity ratio, %;

- The coefficient of concentration of equity (autonomy), %;

- Coefficient of maneuverability of own funds, %;

- The coefficient of availability of own working capital, %;

- The coefficient of concentration of borrowed capital, %;

- Current liquidity ratio, %;

- Quick liquidity ratio, %.

According to these indicators, the data were checked for emissions according to the three sigma rule and the corresponding lines were deleted. For example, emissions from the tax burden for the extractive industry are shown in Figure 2.

Figure 2 – Graphs of emission diagnostics in the source data

Source: compiled by the author according to SPARK-Interfax. ^[12]

For the purified data, machine learning models were built to predict the tax burden based on decision trees (decision tree, gradient boosting), the nearest neighbor method, taking into account the nonlinear nature of the relationship, as well as classical econometric models based on linear regression (linear regression). All models were built "with a teacher", i.e. the dependence itself was determined by a training sample, which included 80% of the initial data, and the diagnosis of the model was carried out using test data (20% of the initial total). Thus, it is possible to obtain more reliable research results by checking the quality of models on independent variables.

Table 4 – Assessment of the quality of models built for different sectors of the economy

Source: compiled by the author according to SPARK-Interfax ^[12]

According to the table, it can be seen that after the data preprocessing stage, from 50 to 76% of the original sample was lost. The linear model based on regression showed worse quality than the methods based on decision trees and the nearest neighbor method.

To assess the quality of the model, the coefficient of determination was used, characterizing the proportion of variation of the dependent variable from the factors included in the model. According to this coefficient, the quality of the gradient boosting model can be distinguished, which showed the best results for all the industries under consideration. At the same time, the highest quality index was 62.1% for manufacturing industries, i.e. 62.1% of the variation in the tax burden is explained by the factors included in the model. This suggests that the industry is sensitive to tax changes, which requires a substantial assessment of all possible consequences of tax transformations. Also, a fairly high quality of models was obtained in the construction and mining industries (determination coefficients of 45.2 and 40%, respectively). For these industries, it can also be concluded that there is a high tax sensitivity. In wholesale and retail trade; repair of motor vehicles and motorcycles, dependence was found at the level of 35.3%. For this industry, the tax sensitivity is low. There is no relationship between economic and financial indicators in financial and insurance organizations (the maximum coefficient of determination is 7.9%). Thus, this industry does not depend on the tax burden in any way and any changes in the field of taxes should not significantly affect the activities of organizations in this industry.

Thus, it can be concluded that industries such as mining, manufacturing and construction in general are quite sensitive to the tax burden. Despite the fact that organizations of wholesale and retail trade; repair of motor vehicles and motorcycles are less dependent, and organizations of financial and insurance activities do not depend on the tax burden in any way.

Conclusions and suggestions

The indicator of the tax burden of industries is important for assessing economic efficiency and investment attractiveness. At the same time, the tax burden varies greatly by industry, depending on regional conditions, as well as on government regulation. Depending on the specifics of a particular industry, for example, due to specialized taxes and duties in the extractive industry, the tax burden increases significantly. Also, the tax burden can be significantly reduced due to special tax benefits, for example, in the IT sector. Also, the reaction of each industry to tax changes varies greatly, which is confirmed by the models built in the work.

The greatest tax sensitivity was manifested in the manufacturing industry. These conclusions coincide with the results of the study by E. V. Balatsky and N. A. Ekimova, despite the fact that their study used macroeconomic indicators in dynamics and models were selected mathematically. For this industry, any tax changes can significantly affect the level of production. At the same time, an additional tightening of the tax regime for the industry may well provoke a decline in production ^[1]. Similar conclusions can be drawn in relation to the construction industry.

As for the extractive industry, which also has a fairly high level of tax sensitivity, it is necessary to take into account its features, which consist in the presence of large companies on the market to a greater extent, as well as in minimizing the use of raw materials, unlike manufacturing and construction, therefore tax changes to a lesser extent can lead to a decline in production, due to this, the tax burden of this industry is the highest among those considered (65.6%, as opposed to 33.5% for manufacturing in 2023).

As for the industry "wholesale and retail trade; repair of motor vehicles and motorcycles" with moderate tax sensitivity. This industry is characterized by high profitability due to low costs (no purchase of expensive equipment, machinery, etc. is required, unlike in production areas) and fast payback. However, the industry is characterized by high risks: high competition, supply problems, low-quality goods, falling demand, etc. The tax burden of this industry is quite high and amounted to 23.7% in 2023. Therefore, it can be concluded that the introduction of tax instruments will not have such a strong impact on large companies in the industry, but for small and medium-sized firms, changes can significantly affect income levels. A number of papers also note that the impact of tax instruments is more significant on small firms than on large ones ^{[18, 19]}. Therefore, for this industry, it is necessary to differentiate tax instruments depending on the scale of activity.

Financial and insurance activities are characterized by low tax sensitivity, therefore, it can be concluded that the introduction of new tax instruments will not have a significant impact on the activities of the industry.

In general, it is necessary to apply tax instruments taking into account the tax sensitivity of each industry and apply the tools in a differentiated manner. Special tax incentive instruments, due to their purposefulness and selectivity, can be effective in solving the problem of accelerating the economic growth of individual industries ^[20]. At the moment, this rule is valid on a limited scale, which requires deepening the methods depending on the specifics and technological level of the industries. At the same time, these tools can be implemented only when determining the optimal conditions between industry representatives and government agencies.