Intel Hub

UMDAC Datathon 2026 · Universiti Malaya · compiled from champions' actual repos, real submission decks and verified problem statements.

Registration closes

—

Nov 9 · 23:59 SGT

Dataset drops

—

Nov 13 · 11:40 SGT

RM 6,000

champion prize

200 → 30

teams → finalists (2025)

public team repos archived

2/3

champions identified

01Key dates

Mon Nov 9
Registration closes action
Team of 1–3 locked at submission · student ID + resume per member (resumes go to sponsors)
Tue–Thu Nov 10–12
Online workshops
Optional, hands-on — judge expectations and sponsor-domain hints
Fri Nov 13 · 11:40
Opening + dataset & problem statement release go
Sponsor session reveals the domain · preliminary round begins
Mon Nov 16
Preliminary deadline submit
Slides ≤10p + dashboard ≤5p (PDF) · no exact cutoff published — verify in Discord
Mon Nov 23
Finalists announced (top 30)
Only teams attending the physical final are considered
Tue Nov 24
Finalist briefing (online)
Time TBC via Discord
Fri Nov 27 · 23:59
Final submission deadline hard
Slides ≤15p + dashboard ≤5p · implementation plan + KPIs expected
Sat Nov 28
Physical final + closing @ FSKTM, UMKL (DK1/BK1)
15-min closed-door pitch · meals provided · business casual

02Winner board

2025 · UMDAC × AstraZeneca

🥇 Later Eat What — led by Zi Yin Shee (UM Statistics)
🥈 Gunners · 🥉 Winnie the Data Scientist
Problem: corporate cash-flow forecasting on AZ data (8 countries)
Champion stack: Python + R + Power BI · 56+ hrs · beat 200 teams
Aftermath: Data Scientist job @ AstraZeneca
LinkedIn → · study the 2025 dashboard in the Library →

2024 · Cryptocurrency × Balaena Quant

🥇 Hokkien Mee Is Red — Sim Sze Yu, John Ong Ming Hom, Lei Wing Teng (UM)
🥉 Milo Ikat Bawah (USM) · consolation Central Limit Legends (UPM)
Winning model: time-series XGBoost → buy/hold/sell thresholds
Backtest 227.9% / Sharpe 1.66 → honest forward Sharpe −1.00
Champion's final slides (PDF) →
study the deck page-by-page in the Library →
github.com/szeyu →

2023 · Multi-theme (DOSM open data)

🥇 Goofy Goober — never publicly identified
🥈 OnzLa · 🥉 Sri Aman Power
Theme 3 "Economy Price Catcher": DOSM household prices
Top-10 team dotdotdot: 3 hypotheses → EDA → Prophet → recs
dotdotdot submission (PDF) → · in the Library →

03Champion playbook

Model → decision → actionNever stop at insights. XGBoost emitted buy/hold/sell; Prophet drove a liquidity plan. Hand judges a decision system.
Earn every choice with numbersModel bake-off (Moving Avg $4.80M vs ExpSmooth $3.72M vs Prophet $2.24M MAE), feature-importance elbow (n=12), accuracy-by-horizon.
Validate multiple ways, report honestlyForward split + rolling windows + monthly CV. The 2024 champion showed a negative forward Sharpe and still won.
Pitch a strategic plan for the sponsor, by name"Balaena Quant can teach clients…" — recommendations prioritized by impact, framed for the sponsor's business.
Simple stack, deep rigorpandas + XGBoost/Prophet + Power BI/Looker. No deep learning anywhere in any winning material.
Personality landsMemorable team names, one honest-limitations slide, humor in the closer. Judges are humans after 30 pitches.

04Problem statements (verified)

2023 — public open data

Household item prices across states/premises (DOSM PriceCatcher)
Deliverable pattern: hypotheses → EDA → forecast → business recs

2024 — crypto markets

"Understanding Cryptocurrency Price Dynamics and Developing Profitable Trading Strategies"
Hourly BTC data 2020–2024, 43 on-chain/market columns · data dictionary

2025 — corporate finance

AstraZeneca multi-entity cash flow: forecast 26 weeks, explain AR/AP, detect anomalies, liquidity risk
Sponsor = data source; dashboard in sponsor's brand colors suggested

Meta-pattern: sponsor industry ≈ problem domain. Every year reduces to understand drivers → forecast → anomaly/risk → recommend to sponsor by name. Watch the 2026 sponsor announcement. · open the Library to study the evidence →

05Sprint A — champion replication (executed)

The 2024 winner's pipeline rebuilt and run on the real dataset: 41,498 hourly rows → 33 features → XGBoost → threshold sweep → backtest with fees.

+7.54%

test return

0.37

test Sharpe (ann.)

−11.17%

max drawdown

828

trades

17.67

train Sharpe → collapse

+93.67%

buy & hold same period

Backtest equity curves — backtest_equity.png — strategy vs buy & hold

Feature importance — feature_importance.png — top 20 XGBoost features

Elbow curve — elbow_curve.png — champion's n=12 marked

Threshold sweep — threshold_sweep.png — Sharpe vs |threshold|

The lesson: train Sharpe 17.67 → test 0.37 reproduces exactly what the champions showed (1.66 → −1.00). Rigorous pipeline + honest validation + a decision layer beats a polished fake. Full code: sprint_a_pipeline.py · metrics: metrics.json · feel it in the Threshold Explorer →

06Prep checklist

Lock team (1–3) — unchangeable after registering
Polish resume (shared with industry sponsors)
Join official Discord once confirmed (mandatory)
Watch sponsor announcements — predict the domain
Drill: Prophet forecasting + MAE/RMSE/MAPE + CIs
Drill: XGBoost sliding windows + SHAP + elbow curve
Drill: anomaly detection (spikes + duplicates)
Drill: dashboard — KPI cards, slicers, forecast overlay
Sprint B: cash-flow drill on 2025 AZ data
Pre-build deck skeleton (in /templates)
Memorable team name (champion precedent exists)
Attend both workshops Nov 10–12