Save taxes without breaking laws using ML
Novel use cases for ML in the
accounting and tax space
Intro
• Chartered Accountant and former tax analytics director
at EY
• Co-founder at a data sciences solutions startup
• Currently building GST solutions with data sciences
Why ML in taxes?
• Finding tax savings options without breaking the law is
the holy grail in tax consulting. There are differences in
impact of taxes, therefore opportunities, for e.g
– The effective tax rate for a company making a profit up to Rs 1
crore was 30.26% in 2015-16 while the corporate tax rate was
25.90% for those with profits greater than Rs 500 crore
– Drugs and pharmaceuticals paid tax at 24.2%, electronics paid
tax at 35.5%
Source: Indiaspend.com
Tax laws are not really English
The major input variables are not simple:
A. Laws are unstructured text data and are continuously changing
B. Legalese is not exactly English
C.Financial statements are a mix of structured and unstructured
information
D.The level of precision required in tax decisions is high
Our method to the madness
• Tokenization
• Stemmers / Lemmatizers
• String distance algorithms
• LSTM models
Combined with
• Tax and accounting domain knowledge
Our key learnings
• Domain knowledge trumps ML results in ultimate
decision making
• Kaggle style ensembling and blackbox solutions are
much harder to sell
• Feedback and reinforcement learning is critical to
achieve higher accuracy
• Feature engineering is very important
Thankyou
Twitter: gsj2
Email: jay@vyakya.com
Linkedin: https://www.linkedin.com/in/jayendrangs

Machine learning in taxation

  • 1.
    Save taxes withoutbreaking laws using ML Novel use cases for ML in the accounting and tax space
  • 2.
    Intro • Chartered Accountantand former tax analytics director at EY • Co-founder at a data sciences solutions startup • Currently building GST solutions with data sciences
  • 3.
    Why ML intaxes? • Finding tax savings options without breaking the law is the holy grail in tax consulting. There are differences in impact of taxes, therefore opportunities, for e.g – The effective tax rate for a company making a profit up to Rs 1 crore was 30.26% in 2015-16 while the corporate tax rate was 25.90% for those with profits greater than Rs 500 crore – Drugs and pharmaceuticals paid tax at 24.2%, electronics paid tax at 35.5% Source: Indiaspend.com
  • 4.
    Tax laws arenot really English The major input variables are not simple: A. Laws are unstructured text data and are continuously changing B. Legalese is not exactly English C.Financial statements are a mix of structured and unstructured information D.The level of precision required in tax decisions is high
  • 5.
    Our method tothe madness • Tokenization • Stemmers / Lemmatizers • String distance algorithms • LSTM models Combined with • Tax and accounting domain knowledge
  • 6.
    Our key learnings •Domain knowledge trumps ML results in ultimate decision making • Kaggle style ensembling and blackbox solutions are much harder to sell • Feedback and reinforcement learning is critical to achieve higher accuracy • Feature engineering is very important
  • 7.
    Thankyou Twitter: gsj2 Email: jay@vyakya.com Linkedin:https://www.linkedin.com/in/jayendrangs