Where’s value? Using VAT data to improve compliance

  • In India, the low Value Added Tax (VAT) compliance of firms has been a persistent problem in raising public revenue.
  • With the use of a large dataset – the universe of VAT transactions in the state of Delhi over the last five years — this project addressed the issue of Indian firms’ evasion of VAT.
  • Using the dataset, researchers evaluated the impact of an experimental policy that aimed to increase VAT compliance by increasing the obligatory paper trails that firms had to keep for reporting to the tax authority.
  • Researchers also developed a machine learning tool that identifies potentially ‘bogus’ firms, which audits should focus on.

The state’s ability to effectively tax its firms is central to a country’s economic development. Tax compliance by firms in India remains low. When it comes to Value Added Tax (VAT), Indian tax officials face the joint challenges of encouraging greater participation and accuracy in reporting by all registered firms, and identifying ‘bogus’ firms that submit fraudulent receipts for genuine ones.

In collaboration with the central tax authorities of the state of Delhi, researchers used a large-scale dataset – the universe of VAT statements in the state of Delhi from the past five years – to evaluate how tax particpation and revenue can be improved.

The dataset allowed an insight into the effects of a policy introduced in the state Delhi in 2012. Although VAT was introduced in Delhi in 2005, until 2012-13 firms were only required to file a single aggregated tax return (known as a consolidated return). The consolidated return did not contain enough information to verify buyer reports and seller reports – this would have to be done through an audit.

In 2012, tax authorities in Delhi introduced a policy mandating firms to file detailed information about their transactions with other registered firms. With this new information, the tax authority could now relatively easily cross-check information provided by registered buyers with the information from corresponding registered sellers directly on its own servers without initiating a potentially costly audit. In case of a mismatch between buyer and seller reports, automatic notices are now sent out to both firms who are then required to amend their respective returns so that they are in agreement.

As expected, the policy led to a significant increase in tax collected from large wholesale firms who tend to trade with other tax registered firms, but not as much from retail firms who will tend to trade with unregistered firms and individuals.

As well as this policy evaluation, researchers also used machine learning techniques to develop a predictive algorithm that could identify ‘bogus’ firms based on a training data set with identified fraudulent firms. These firms were flagged for potential audit by tax authorities. Researchers trained local tax officials in using these machine learning techniques to further develop this approach.

Both the lessons of the policy and the newly developed machine learning techniques are now being utilised in Tamil Nadu as well.

Outputs