Aryan Meena

Data scientist and analyst, Boston Updated 26 October 2026

Aryan Meena

Good data first. Everything else comes after.

I got hooked on data at an ISRO internship, traded FX and commodities on prop-firm challenges, and came to Boston to learn how quant and risk teams test strategies. Checking the data first is how I found that a 15-year pattern in the U.S. Dollar Index came from bad prices in a free feed. That paper is now published in Risks.

Where I’m headedQuant research, trading analytics and risk, at banks, quant firms and asset managers.

Just as ready forData analyst, data engineer, data scientist and research roles, in any industry.

  • First-author papers, two published in October 20265
  • Prop-firm profit targets hit in a row3
  • Loss-limit breaches while trading0
  • Wikipedia edits modelled on a Spark cluster60M
  • Live apps and dashboards you can open today5
Read my story See the work

Ask me anything

Answers come from notes I wrote and checked.

Hiring for

Hi, I’m Aryan. Pick what you’re hiring for and I’ll point you to the right work, or ask me anything: a project, a tool, my visa status, why I’d fit your team.

Looking for
Quant, risk and trading analytics first; also data analyst, data engineer, data scientist and research roles. Entry level, anywhere in the US.
Education
MS Applied Data Analytics, Boston University, May 2026, GPA 3.75. BTech, IIST.
Location and start
Boston, open to relocating anywhere in the US. Available now, full time.
Work status
Authorized to work in the US (F-1 STEM OPT). I’ll need H-1B sponsorship later.
Résumé and links

LinkedIn, GitHub, ORCID

The same overnight bet on the dollar, on the same 3,793 nights, measured two ways. Which one would you trade?

Figure 1. $100 held only overnight (previous close to next open) in the U.S. Dollar Index, December 2010 to December 2025. On the free feed of the spot index, the bet loses money year after year, Sharpe −1.36, and standard overfitting tests rated that result robust (PBO 0.02 vs 0.77). On the ICE Dollar Index futures contract over the same nights, it ends slightly up, Sharpe +0.25, not statistically different from zero. Recomputed for this page from the paper’s public data files.

The data behind it

SeriesWhat it isSourceDays
U.S. Dollar Index, spotThe ICE index, calculated from six exchange rates. It can’t be traded directly.Free public feed3,932
Dollar Index futures (DX)The exchange-traded ICE contract on the same indexBarchart, exchange-sourced3,889
Currency futuresCME euro, sterling, Canadian dollar, yen and Swiss franc contractsBarchart, exchange-sourced3,800
Spot rates and interest ratesFederal Reserve H.10 noon rates; OECD interbank ratesFRED6,929

“But spot and futures aren’t the same thing.”

Right, and that’s why futures alone couldn’t settle it. They can differ in trading hours, in how a daily bar is cut, and in carry. So the paper ran three more checks:

  1. Spot against spot. I rebuilt the index from the free feed’s own six currency prices with ICE’s published formula. Its overnight Sharpe is +0.22, close to futures and nowhere near −1.36.
  2. Carry. If the futures premium explained the gap, it would move the overnight and daytime legs the same way. Instead the gap opens overnight (−0.014 index points) and closes by the next close (+0.013). Only the open is off.
  3. The open itself. The free feed recorded the open at the day’s exact low on 204 days and at the high on 82. On those days it sat about 40 basis points from the rebuilt spot open.

The closes agree: Federal Reserve noon rates from FRED track both series’ daily moves (correlation 0.86). The error lives in the opening prices.

§1 How I got here

The short version is on my résumé. This is the real one.

  1. 2019 – 2023

    It started with labels

    My first real taste of data was an internship at ISRO’s Space Applications Centre, training models to find roads in satellite images. The model wasn’t the lesson. The labels were. Something as simple as how an image is labelled decided whether the whole thing worked.

    My thesis on meta-learning (MAML), and all the computer vision and machine learning that came with my degree at IIST, kept teaching me the same thing: understand the data, clean it, and squeeze every bit of information out of it before you run a single model on it.

  2. 2023 – 2024

    The hard year

    I wanted depth, so I wanted a master’s. And I wanted the US: the tech hub, and a shot at working for the companies I’d heard about my whole life.

    The year I spent applying was the hardest one I’ve had. Things were tough at home, and the job offers weren’t coming. So I didn’t wait. I published two research papers to make my profile stronger, and I taught myself to trade.

  3. Dec 2023 – Dec 2024

    Learning the market by trading it

    I traded EUR/USD, USD/JPY, USD/CAD, gold, silver and oil. Every day I read the higher timeframes, then executed on the lower ones, mostly around the London and New York opens.

    Every strategy went through the same loop: backtest it, trade it live, collect the data, then look hard at win rate, maximum drawdown, what failed and how often I passed. I kept adjusting my execution, my analysis and how I read market and news sentiment until the numbers held. That’s how I passed prop-firm challenges at Funding Pips, hitting a 15% profit target three times in a row without ever breaking the 10% total or 5% daily loss limit.

  4. 2025 – 2026

    Boston, and learning it from the base

    At Boston University I wanted to learn how it’s really done: how quant researchers and traders at banks and quant firms build strategies, and what model and risk analysts check before any money moves. With Prof. Eugene Pinsky I went back to the markets I’d traded, first on free data, then on exchange vendor data, then on Bloomberg.

    I turned the math into code and learned the market’s realities the honest way. A great Sharpe ratio in one period can fall apart in the next. A Newey-West t-test, a deflated Sharpe ratio and a maximum drawdown each tell you something different. And which data vendor you trust can change the answer completely. On a free feed, a 15-year pattern in the U.S. Dollar Index looked like money. On exchange data it disappeared. Proving why became my first-author paper in Risks.

    Research was only half of it. My coursework covered machine learning, big data analytics, data mining, algorithms, graph analytics and data visualization, and the class projects grew into real work. WikiFlow started as my big data term project and was published by Springer. My data science class project became the readmission study. A team project on graph partitioning became the MBTA resilience simulator.

    Along the way I taught machine learning to 60+ graduate students as a teaching assistant and led a six-person team to a podium finish at the BU MedAI Hackathon.

  5. Now

    What I want to do next

    My first choice is finance: quant research, trading analytics or risk at a bank, a quant firm or an asset manager, where data quality and honest testing decide real money.

    The same habits work anywhere, though, and I’d be just as glad to join any company as a data analyst, data engineer, data scientist or researcher. Either way, I bring work that already runs on its own: an AWS pipeline that won’t publish until 17 checks pass, a hospital call list that reaches 3× more readmissions than random calls, and an app that catches AI agents reaching the right answer the wrong way.

What I bring to your team

  • Market sense. I’ve traded FX and commodities with real stakes, so drawdown, costs and risk limits aren’t theory to me. In my research I price every strategy in basis points before I call it a result.
  • Data first. I clean and check before I model. My screen caught 10 faulty price feeds out of 37 before they reached analysis, and my AWS pipeline refuses to publish until 17 integrity checks pass.
  • Work that runs. Live right now: an AWS dashboard, a readmission dashboard, a Power BI report, a subway resilience simulator and an AI evaluation app. Five first-author papers behind them.
  • Plain English. I taught 60+ graduate students machine learning, and I build every dashboard so someone without code can check the numbers.

§2 Selected work

Every project in four parts: the problem, what I did, the tools, and the result. The live figures run on the real exported data, so go ahead and play with them.

Does a fifteen-year pattern in the dollar survive a second data source?

Two blind spots in backtest validation

Problem
A strategy on the U.S. Dollar Index showed a 15-year overnight pattern, and standard overfitting tests rated it robust. Before anyone puts money on a result like that, someone has to find out whether it’s real.
What I did
Rebuilt the test on independently sourced data: ICE Dollar Index futures and five CME currency futures from Barchart, plus Federal Reserve spot and interest rates from FRED. Rebuilt the spot index from its six currencies with ICE’s formula. Wrote a one-line screen that flags bad opening prices, and a search engine that reproduces the original backtest exactly on all 1,312 configurations, about 17× faster, so the whole search could be rerun 2,000 times against luck. Research assistantship with Prof. Eugene Pinsky.
Tools
Python (pandas, NumPy, statsmodels), Excel for the first pass, SQL for the published dashboard. Newey-West, White’s Reality Check, Hansen’s SPA, CSCV.
Result
The pattern lived only in the free feed’s opening prices: −1.36 there, +0.25 on futures and +0.22 on the rebuilt spot index. Published as first author in Risks (October 2026). The same screen flagged 10 faulty series out of 37 price feeds before they reached any analysis.
−1.36 → +0.25overnight Sharpe, free spot feed vs futures
10 of 37price feeds flagged before use
17×faster search, identical results

Who should a care team call first after a diabetic patient goes home?

Readmission risk, re-analysed and put to work

Problem
About 1 in 9 diabetic patients in this data came back to hospital within 30 days. Care teams can’t call everyone, and a published 2026 benchmark set the bar for predicting who will return at AUC 0.664.
What I did
Re-ran the benchmark under the same nested, patient-grouped cross-validation; rebuilt the patient cohort and added patient-history features; tested two Transformer models against gradient boosting; audited accuracy by age, race and sex; turned the predictions into a call-list dashboard a care team can size to its capacity. It started as my class project in Data Science with Python; I later rebuilt it from scratch.
Tools
Python, LightGBM, XGBoost, scikit-learn, Tableau, HTML and JavaScript dashboard.
Result
AUC 0.689 against the published 0.664 on 101,766 encounters. Calling the riskiest 5% reaches 1,696 future readmissions, 3.0× what random calls reach. The simpler model shipped because the Transformers lost at up to 65 times the compute. The audit found no gap by race or sex, but a drop from 0.73 to 0.63 across age bands, which argues for age-aware thresholds.
Calling the riskiest 5% reaches 1,696 future readmissions, 3.0× the 566 that random calls reach. Source

Can a reporting pipeline refuse to publish a wrong number on its own?

Service operations analytics on AWS

Problem
Reporting pipelines fail quietly: duplicate records, late corrections, and a bad run that overwrites yesterday’s good report.
What I did
Designed an AWS pipeline that replays 105,372 incident deliveries, handles duplicates, late arrivals and corrections, and refuses to publish unless 17 integrity checks pass. Added a daily NYC 311 demand forecast for the five boroughs with written rules before any new model can replace the current one, and a resolution-time model that keeps unfinished tickets in the data.
Tools
AWS S3, Step Functions, Glue, PySpark, Athena, Lambda, CloudFront, GitHub Actions with a restricted OIDC role, pytest.
Result
Live dashboard. A forced failure left the last good report untouched, and a repeated request finished without rerunning compute. About $0.03 per processing run, guarded by 81 automated tests.
  1. S3 replay batches
  2. Step Functions
  3. Glue, Python and Spark SQL
  4. 17 integrity gates
  5. Parquet facts and dimensions
  6. Athena verification
  7. CloudFront dashboard

Can you tell who’ll quit Wikipedia from their first ten edits?

WikiFlow: editor dropout at scale

Problem
Wikipedia loses most new editors; in the test set, 64.8% dropped out. Mentors need to know whom to reach first, from very little history.
What I did
Processed five years of English Wikipedia history on a 5-node Google Cloud Dataproc cluster. Engineered 13 features from each editor’s first ten edits and threw out co-editor graph features because they leaked the answer. Wrote logistic regression from scratch on Spark RDDs, distributed a Keras network with Elephas, and clustered newcomers into types. Built as my Big Data Analytics term project with Kunj Patel, then published with Prof. Farshid Alizadeh-Shabdiz.
Tools
PySpark, GCP Dataproc, BigQuery, Keras with Elephas.
Result
About 60 million revision events from 1,158,248 editors (25 GB compressed) for roughly $2.50 a full run. AUC 0.909 with the neural net, 0.845 in 19 seconds with the from-scratch model. Four kinds of newcomer. Published by Springer.

Four kinds of newcomer the clustering found

Test set of 231,615 editors; 64.8% dropped out. Source

Which single closure hurts Boston’s subway most?

MBTA network resilience

Problem
Which subway stations, if closed, cut Boston’s network apart? The public schedule data had defects that dropped the Green Line and broke the map into six pieces.
What I did
Rebuilt the MBTA’s Spring 2025 schedule (20.7K trips, 355K stop-time records) into one connected network after fixing three pipeline defects. Compared Louvain, spectral and METIS partitioning across 25 seeds, simulated every closure, and shipped an interactive simulator. It began as a three-person class project that I later rebuilt and extended.
Tools
Python, NetworkX, JavaScript, GitHub Actions.
Result
125 stations and 127 segments with only three independent loops, so 110 of 125 stations are single points of failure. Closing Copley alone cuts off 56.1% of station pairs. Try it on the map: the numbers are computed live in your browser.
56.1% of station pairs can no longer reach each other with Copley closed.

Can you trust an AI judge to grade an AI agent?

JudgeGuard: auditing agent traces and LLM judges

Problem
Teams use AI models to grade AI agents. But an agent can reach the right final answer through a wrong step, and an AI judge can score a wrong answer as highly as a right one.
What I did
Built an evaluation workbench that replays an agent’s tool calls and flags the exact step that doesn’t add up, and tests AI judges against source-supported answers with a real local model. Shipped it with continuous integration, property-based tests and a hardened Docker image.
Tools
Python, FastAPI, Gradio on Hugging Face ZeroGPU, Qwen3, Docker, GitHub Actions, Hypothesis.
Result
Live on Hugging Face. A Qwen3-4B judge ranked the supported answer first on 96 of 100 pairs (95% interval 92.45–99.03%), with all 200 raw responses published. The guard path meets a 150 ms p99 latency budget. Try the demo’s opening case on the right.
  1. lookup_units() → 4200
  2. lookup_price() → 28.5
  3. multiply(4200, 28.5) → 115500
  4. final_answer “$119,700”

Steps 1 and 2 are illustrative; the multiplication is the demo’s own example.

Should commodity investors chase last month’s winners?

Commodity rotation in futures and ETFs

Problem
Should commodity investors rotate into last period’s winners, and does the answer change between futures and ETFs?
What I did
Tested winner, median and loser rotation on the 15 most liquid Bloomberg commodity futures (2008–2026) and six commodity ETFs, rebalancing weekly to yearly, against equal weight, random portfolios, BCOM and DBC. Corrected for the number of strategies tried and measured each one’s break-even trading cost. With Prof. Eugene Pinsky.
Tools
Python, Bloomberg total-return indices, random-portfolio benchmarks, White’s Reality Check, Hansen’s SPA.
Result
Same rule, opposite answers. In ETFs, monthly winners turned $100 into $337, against $121 for equal weight and $86 for DBC. In futures, winners trailed equal weight at every horizon from a month to a year. Weekly rotation breaks even at only 4–5 basis points. Under review at Commodities.
Growth of $100. ETFs 2008–2025; futures Jan 2008 – Sep 2026. Source

How wrong is the bell curve when markets crash?

Multi-asset risk engine

Problem
Many risk models assume returns follow a bell curve. In a crash, that assumption can understate the loss a desk should plan for.
What I did
Wrote a multi-asset risk engine over commodities, the S&P 500, five CME currency futures and dollar-index futures (December 2010 to December 2025). Fitted GARCH and DCC models by maximum likelihood, compared Monte Carlo and block bootstrap, and backtested 8 VaR models the way a bank would.
Tools
Python, SciPy, GARCH and DCC, Monte Carlo, Kupiec and Christoffersen backtests, pytest.
Result
The normal model understated one-month 99% expected shortfall by about a third (11.1% vs 7.4%). Over 2,796 days, only DCC-GARCH with an empirical tail passed the full conditional-coverage test, while holding about 13% less VaR than historical simulation. On a 40/30/30 book, 52% of the risk came from the 40% equity weight.
One-month (21-day) 99% expected shortfall on the 40/30/30 book. Source
Code

What actually moves review scores in an online marketplace?

RetailPulse: SQL and Power BI

Problem
An online marketplace with about 100K orders across nine messy tables. What actually drives reviews and revenue?
What I did
Cleaned the tables into a PostgreSQL star schema (raw, clean, mart), wrote cohort-retention and RFM segmentation in SQL, and published a four-page Power BI Service report with DAX measures.
Tools
PostgreSQL, SQL (CTEs, window functions), Power BI Service, DAX.
Result
Late deliveries drop the average review from 4.29 to 2.57 stars. Only 3.0% of customers ever order again, and the top two RFM segments are 40% of customers but 74% of revenue: fix delivery and spend on acquisition before loyalty programs. 98,199 orders, R$15.74M in sales after cleaning.
Average review score by delivery outcome. Source

Can one 3D model read brain scans from four different PET tracers?

Amyloid PET Centiloid prediction

Problem
Hackathon challenge: predict amyloid burden, a marker used in Alzheimer’s research, from 3D brain PET scans taken with four different tracers.
What I did
Led a six-person team as technical lead. Used one 3D MedicalNet ResNet-34 backbone for all four tracers, with FiLM conditioning after every residual stage, mixed precision and a Huber loss.
Tools
PyTorch, MedicalNet, FiLM conditioning.
Result
Podium finish among 40+ teams at the BU MedAI Hackathon (Spring 2026). Validation MAE 7.31 Centiloid with Pearson r 0.968 on 500 scans.
Code

How often do AI assistants get NASA’s exoplanet archive right?

ExoQuery-Bench In progress

Problem
AI assistants now answer questions over NASA’s Exoplanet Archive by writing database queries. There’s no public, reproducible score for how often they get it right.
What I did
Building an open benchmark that scores Claude, OpenAI, Gemini and a local model across plain prompts, RAG, a self-correcting agent and the AstroFetch tools, and reports where answers fail. Logged the archive’s traps first.
Tools
Python, ADQL, RAG, agents, MCP tools.
Result
In progress. Traps found so far: a naive count of transit planets returns 36,100 rows instead of 4,711 planets, and mixing Jupiter and Earth radii is an 11.2× error. No model scores yet; they go up when they’re real.

Also in the notebook

StudyResultTools
Driver drowsiness detection (Advanced Machine Learning project)ROC-AUC 0.964, F1 0.923 on NTHU-DDDPyTorch, RetinaFace, ResNet-18, random forest
Learning-difficulty screening, NHIS 202436 models compared on 7,439 children and 353 variablesPython, XGBoost, imbalanced-learn
Jet-engine failure warning, NASA C-MAPSSAUC 0.945 for a 30-cycle warningR, lasso logistic regression, ANOVA
Vehicle-to-everything network performanceAdjusted R² 0.869 on 13,208 observationsR, polynomial regression
Road segmentation at SAC/ISROU-Net and DeepLabV3+ on 6,723 aerial imagesPyTorch, QGIS
BTech thesis, meta-learning for blind super-resolutionPSNR 33.513, SSIM 0.9324 on Set5PyTorch, MAML

§3 Papers

Five as first author. Two came out in the same week of October 2026.

  1. Published23 Oct 2026

    Meena, A., & Pinsky, E. (2026). Two Blind Spots in Backtest Validation: From a Manufactured Dollar Index Anomaly to a Corrected Null Result. Risks, 14(11), 247.

    Two ways a backtest can pass every standard check and still be wrong: bad input data, and a search nobody declared. Figure 1 is from this paper.

    doi:10.3390/risks14110247 Code and data

  2. Published19 Oct 2026

    Meena, A., Patel, K. M. K., & Alizadeh-Shabdiz, F. (2026). WikiFlow: Predicting Wikipedia Editor Dropout at Scale Using Distributed Machine Learning on Google Cloud. In Computer Science and Education in Computer Science (CSECS 2026), LNICST vol. 725, pp. 361–376. Springer.

    Which new Wikipedia editors will quit, predicted from their first ten edits, across 1.16 million editors on a Spark cluster.

    doi:10.1007/978-3-032-42320-7_22 Code

  3. Under reviewSubmitted Oct 2026

    Meena, A., & Pinsky, E. Should Commodity Investors Chase the Winners? Momentum, Median, and Contrarian Rotation in Commodity Futures and ETFs. Under review at Commodities (MDPI).

    The same rotation rule gives opposite answers in commodity futures and commodity ETFs. Read the summary.

  4. Published2024

    Meena, A. (first author) (2024). Accelerated CNN Training with Genetic Algorithm Optimization. IEEE IATMSI 2024.

    A genetic algorithm searches CNN training settings so a convolutional network trains in less time.

  5. Published2024

    Meena, A. (first author) (2024). Enhanced Scene Text Recognition for Improved Geocoding and Accessibility. ICADIE 2024, Springer.

    Reading text in street scenes well enough to place it on a map.

Two more manuscripts on commodity futures strategies with Prof. Eugene Pinsky are in preparation. Full list on ORCID 0009-0005-6739-8265.

§4 Experience

Newest first.

  1. Oct 2026

    Two papers published in one week

    Risks (MDPI) on 23 October and Springer LNICST on 19 October.

  2. May 2026

    MS, Applied Data Analytics, Boston University

    GPA 3.75. Coursework: Foundations of Machine Learning, Advanced Machine Learning, Big Data Analytics, Data Mining, Analysis of Algorithms, Web Mining and Graph Analytics, Data Science with Python, Data Analytics and Visualization. Class projects that became real work: WikiFlow (Big Data Analytics, published by Springer), the readmission study (Data Science with Python), the MBTA network study (team project) and driver drowsiness detection (Advanced Machine Learning).

  3. Jan – May 2026

    Graduate Teaching Assistant, Boston University

    Advanced Machine Learning and Data Science with Python, 60+ graduate students: weekly office hours and written feedback on classification, regression, clustering and deep learning projects.

  4. Spring 2026

    Technical lead, BU MedAI Hackathon

    Led a team of six to a podium finish among 40+ teams.

  5. Sep 2025 – May 2026

    Research Assistant, Quantitative Finance, Boston University

    With Prof. Eugene Pinsky. Validated FX and commodity strategies on free, exchange-vendor and Bloomberg data, wrote a multi-asset risk engine, and launched a 12-table SQL dashboard so non-specialists can check every result. Excel for first looks, Python for the heavy analysis, SQL to publish.

  6. Dec 2023 – Dec 2024

    FX Trader, Funding Pips (funded account, remote)

    Traded EUR/USD, USD/JPY, USD/CAD, gold, silver and oil intraday around the London and New York opens. Backtested every strategy, traded it live and tracked win rate and drawdown. Hit a profit target of at least 15% of account size three times in a row while never breaching the 10% maximum or 5% daily drawdown limits.

  7. 2024

    Two first-author papers

    IEEE IATMSI 2024 and ICADIE 2024 (Springer).

  8. Jun – Jul 2022

    ML Research Intern, Space Applications Centre (ISRO), Ahmedabad

    Road segmentation from satellite and aerial imagery with U-Net and DeepLabV3+, with QGIS in the loop.

  9. 2019 – 2023

    BTech, Electronics and Communication Engineering, IIST

    Indian Institute of Space Science and Technology. Coursework in statistics and probability, computer vision and machine learning for signal processing.

§5 Side B

Off the clock I play games, write code for fun and keep music on. These tracks mix all three with what I do at work.

Track 1

Overfit me

Twelve trading strategies, three years of backtest each. Pick the one you’d fund. Then we’ll run the next two years.

Pick a strategy.

Track 2

Rating the raters

Games are my favourite way to switch off, so my side project audits EA FC 26: scraping the player ratings, joining them to real match data, and testing whether the gaps predict next season’s rating changes. Any scoring system can be audited this way, recommenders included. In progress, no findings yet.

Track 3

Hear a backtest

The Listen to it button in Track 1 maps each week of returns to a note on a pentatonic scale: up weeks climb, down weeks fall. A lucky backtest sounds suspiciously like a melody.

Track 4

This page

Plain HTML, CSS and JavaScript with no framework or template, lighter than most single photos. Press / to search it. There’s also a classic cheat code hidden in here, if you grew up on consoles.

References

Every underlined figure on this page points here. Click one to see its source in place.

    ↑↓ to move, Enter to open, Esc to close