Himanshu Mahale
BS Economic Sciences, Indian Institute of Technology, Kanpur
himanshumahale20@gmail.com |
LinkedIn |
github.com/himdev20
Education
Major Hemant Jakate College, Nagpur
2023Vidya Niketan School, Bramhapuri
2021Scholastic achievements
- 99.7th percentile in the Joint Entrance Examination (JEE Main 2023), among 1.2 million+ candidates.
- All India Rank 6763 in JEE Advanced 2023, among 180,000+ candidates.
- Second place among 2,500+ participants in a national technology competition, winning a prize of Rs 50,000.
Research experience
Whose words move rates? RBI MPC minutes and Indian yields
Aug 2026 to present- Scraped all 62 RBI MPC minutes from 2016 to 2026 and parsed them into 372 member statements with votes, building an end-to-end Python pipeline with unit tests that rebuilds every result from raw RBI, FBIL and FRED data.
- Scored each statement’s hawkishness with a name-masked LLM that matches the member’s actual vote in all 372 cases and flags the 40 formal dissenters with an AUC of 0.98, benchmarked against dictionary and VADER scores.
- Separated policy shocks from RBI information effects using Gürkaynak–Sack–Swanson factors and Jarociński–Karadi sign restrictions on OIS and T-bill yields, and found that 42% of decisions carry an information effect.
- Found that a 1-point hawkish gap between members and the Governor lifts the 1-year OIS rate by 2.3 bp on minutes day (p = 0.015), driven by internal RBI members, with null placebos but a weaker effect outside 2020–21.
Effect of trade policy and geopolitical risk on Indian markets
Jan 2026 to Apr 2026- Assembled a 321-month panel of Nifty 50 returns, news-based uncertainty indices and CEIC macro series, ratio-splicing three IIP base years into an HP-filtered output gap and validating every outlier against RBI data.
- Derived and coded a GARCH-X(1,1) likelihood in Python with Hessian-based standard errors and multi-start checks, finding persistence of 0.94 and no joint gain from TPU and GPR in a likelihood-ratio test (p = 0.48).
- Chose ARDL bounds tests, a Cholesky SVAR and a Kalman-filter TVP model from unit-root and Johansen evidence, finding a long-run link via US monetary uncertainty (F = 6.75) but no significant TPU effect in any month.
- Benchmarked LASSO, Ridge, Random Forest and XGBoost against a simple AR(1) that won out of sample (R² 0.79 vs 0.77 and 0.38) and used Diebold–Mariano tests and SHAP to show ML adds no signal beyond persistence.
Rural public-service delivery cost
Sep 2025 to Dec 2025- Replicated a published vehicle routing model in Python for 177 delivery locations and verified that algorithmic optimisation beats judgment-based routing by 15.2%, saving about Rs 100,000 every year in delivery costs.
- Estimated cost-capacity curves under 8, 10 and 12-hour time limits and found that a Rs 20,000 truck retrofit fully pays back in 6.7 months when there is no time limit but saves nothing under a strict 10-hour rule.
- Computed exact Shapley cost shares for all 177 villages and found that the most remote 20% bear 33% of total delivery cost, with a 0.72 rank correlation between each village’s cost and its distance.
- Stress-tested the results against rate, fuel, seasonal and demand shocks and a noisy distance matrix with 57.8% triangle-inequality violations; retrofit payback stays between 4.8 and 9.1 months.
Work experience
Data Analyst Intern, Global Risk and Compliance
May 2026 to Jul 2026- Designed a quasi-experimental test-versus-control framework to measure how third-party fraud changes the spending behaviour of 100K+ American Express customers before, during and after a fraud event.
- Built 12 metrics and 3 new indices in SQL, including a disengagement score that compares each customer’s post-fraud spending with their own pre-fraud baseline, to track churn, dollar loss, spend intensity and recovery.
- Modelled customer recovery timelines with a normalised slope-based method, measuring how quickly spending returns to its pre-fraud level during the 6-month post-fraud window.
- Estimated a ~$37M opportunity cost from fraud-driven disengagement over a 6-month window and identified a portfolio segment with a recurring loss of ~$7.5K every month.
Key projects
Ujjwala: access versus sustained LPG usage in India
Oct 2026 to Nov 2026- Co-authored a policy evaluation of India’s Ujjwala LPG scheme for 2016–2026 using 70+ audit, parliamentary and survey sources, showing that 10.58 crore new connections led to just 4.71 refills per household a year.
- Classified all evidence as administrative, observational or quasi-experimental and showed that the post-2022 rise in refills fits a price effect but cannot be proven causal, given parallel income and enrolment growth.
- Linked low use to free fuelwood, cash-poor households and a connections-only target, as 25–30% of households took 0–1 refills a year and state averages ranged from 1.37 in Chhattisgarh to 5.25 in Haryana.
- Ranked 5 subsidy reforms by weighted multi-criteria scoring, costed 4 subsidy scenarios at an extra Rs 2,900–11,900 crore a year and designed a phased district rollout in which later districts act as the control group.
Monetary and fiscal policy mix in India
Sep 2026 to Oct 2026- Built a reproducible Python pipeline on RBI, MoSPI and Union Budget data from 2004 to 2026 and estimated the output gap with HP, Hamilton and trend filters whose estimates differ by over 6 percentage points in 2025.
- Benchmarked the RBI repo rate against three Taylor rules and an estimated reaction function, finding heavy interest rate smoothing and a 2022 tightening cycle that lagged the rule by 3 percentage points.
- Measured repo rate pass-through to bank lending rates: 82% reaches new loans but only 56% reaches existing loans, while fiscal stimulus hit 3% of GDP in the pandemic.
- Showed that the link between public capital spending and growth fades within two years and is not a causal multiplier, and wrote 15 Pytest tests that replicate every table in the final report.
Applied microeconometric models
Aug 2025 to Nov 2025- Hand-coded Newton–Raphson maximum likelihood in MATLAB for probit, logit and Poisson models, finding that 5 more years of age cut support for legalisation by 1.6 pp and an outbreak elasticity to cattle of 0.95 (t = 5.0).
- Coded Tobit and Heckman selection likelihoods for labour force participation and found significant selection into work (ρ = −0.51, t = −4.5), with bootstrap standard errors 25% higher on the selection term.
- Estimated first-difference and fixed-effects models of US state murder rates and showed that the deterrent effect of executions (−0.093, p < 0.01) loses significance once Texas is dropped (−0.055, p = 0.48).
Relevant coursework
- Economics
- Microeconomics I and II, Macroeconomics I and II, Applied Macroeconomics, Game Theory, Foundations of Financial Risk, Financial Economics, Public Policy in India
- Maths and statistics
- Linear Algebra, Optimization, Several Variable Calculus, Ordinary Differential Equations, Econometrics I and II, Applied Probability and Statistics
- Computing
- Data Structures and Algorithms, Fundamentals of Computing I and II, Computer Vision and Deep Learning, Introduction to Blockchain Technology, Computer Networks
Technical skills
- Languages
- Python, R, Stata, MATLAB, SQL, C, C++
- Libraries
- NumPy, Pandas, Statsmodels, Scikit-learn, SHAP, OR-Tools, Pytest, Matplotlib, PyTorch
- Tools
- Git, GitHub, LaTeX, MS Excel
Leadership
Editor, Vox Populi
May 2025 to Apr 2026- Led a 60+ member student journalism team with 12+ editors handling a Rs 1 lakh budget and directed an 8+ member investigation into academic inbreeding, interviewing 15+ professors including 5+ Heads of Department.
- Pushed for academic and mental health reforms through coverage of the 2024 student protests (50K+ impressions), helping secure a Standard Monitoring Committee, and surveyed 200+ students on hostel issues affecting 1,000+ students.
- Exposed campus drainage flaws, drawing administration attention and raising awareness among 15K+ people on campus, and grew the publication’s readership to 30,000+ in a single quarter.