From the Demo Day event, the best machine learning models were selected, grouped into three categoriesCredit Scoring, ATM Cash Optimization and Fraud Detection.
The Finhacks 2018 #DataChallenge Ends, 14 Data Science Champion Teams Emerge
*the article appears and is taken from Daily Social on 16/11/2018”
The Finhacks 2018 #DataChallenge has reached its pinnacle, namely the final Demo Day. Here, all finalists present their work before the select judges and attendees. The grand finale was held at Soehanna Hall, Jakarta on 14 November, starting from 09.00 until 18.00.
The Finhacks 2018 #DataChallenge is an initiative of PT Bank Central Asia Tbk (BCA), as the best bank in Indonesia and Asia, that aims to help develop the data science ecosystem in Indonesia supported DailySocial.id and Algoritma.
Below is the result of the best team at the Finhacks 2018 #DataChallenge – Demo Day.
Credit Scoring Category
Best Team for Credit Scoring Category
1st Winner – 4.5 Namun Tetap Keren Team, win a prize of Rp80 million
In their presentation, the 4.5 Namun Tetap Keren Team explains that the machine learning model they created is able to facilitate credit scoring system with 86.6% accuracy. The model was created using the stacking ensemble technique, in which Random Forest was utilised as base estimator and boosted with Adaptive Boosting (AdaBoost). From pre-processing data, cleansing, to modelling, all stages can be done in a relatively easy way, utilizing efficient computing capabilities for the banking industry.
2nd Winner – yangmana Team, win a prize of Rp50 million
In their presentation, the yangmana Team explains that the model they designed was a result of their creative work in creating new features, commonly known as feature engineering, supported by XGBoost, a machine learning algorithm, widely known to win the predictive modelling competition. In addition to that, they were also able to explain the model they built using SHAP scoring, where such model is generally hard to explain due to its complex black-box model.
The other two finalists under the same category are FullKeju and growingtaiba. Both teams won a prize of Rp10 million each. Another team, s1mple, was disqualified because of non-attendance at the Demo Day, as specified in the terms and conditions of the competition.
Fraud Detection Category
Best Teams for Fraud Detection Category
1st Winner – exB202 Team, win a prize of Rp80 million
In their presentation, exB202 explains that they designed a simple model and yet offers very high performance, and doesn’t require computers with high computing performance. They utilized Gradient Boosting Decision Trees (GBDT) as core algorithm, with the implementation of XGBoost. They designed special additional features to improve GBDT performance.
Initially, a model without any help feature was built, then the structure was observed. The results of tree structure analysis combined with interaction analysis between variables can produce an additional feature that boosts the model performance. The feature is later called “Duration”. In the end, their model is composed of raw data and one additional feature called “duration”, which then added into XGBoost.
2nd Winner – Pemula Team, win a prize of Rp50 million
In their presentation, the Pemula team explains that they utilize stacking, an ensemble method, as the solution, where it assesses and employs predictions gathered from 10 different machine learning models. Their second machine learning model was also trained to identify mistakes from their previous model. They also use XGBoost, which hyperparameter has been fine-tuned using Bayesian Optimization. Its final predictions is an average of all XGBoost and stacked models to reduce variance and avoid overfitting.
The other three finalists in the Fraud Detection category are 3M, NRGO, and theDoctor. The three teams win a prize of Rp10 million each.
ATM Cash Optimization Category
Best Teams for ATM Cash Optimization Category
1st Winner – dilan Team, win a prize of Rp80 million
In their presentation, dilan team describes that the machine learning they develop is an incorporated model of four XGBoost models, using manifold scenarios of payday variable and 1 Naive model (150 ATMs).
The most important variable to predict money withdrawal in that period is the payday variable, assuming that if a payday falls on a holiday, it is moved back to the previous business day. The predicted results from their model generated the best prediction results compared to the 124 models submitted.
2nd Winner – Resistance Team, win a prize of Rp50 million
In their presentation, the Resistance team explains that they predicted the number of ATM withdrawals at a certain time by utilizing available historical data. When exploring the data, they found that, generally, the number of withdrawals on Sundays and holidays is lower than other days, while at the end of month (near payday), it tends to increase. In their attempt to group ATMs, they found that a number of ATMs are located inside office buildings and shopping centers.
The Gradient Boosting Decision Tree Model they created was trained using the ATM average cash withdrawal patterns. For example, ATM K1 shows a lower withdrawal on Sundays, because the withdrawal amount is only 0.6*mean. Based on the test results on the validation set, the model was able to read the increase and decrease patterns in withdrawals pretty well – compared to their previous time series model.
The other three finalists in the ATM Cash Optimization category are the Nasi Kuning, Talam, Stat-Ion team. All three teams win a prize of Rp10 million each.
“Through the Finhacks 2018 #DataChallenge, we believe that we can find the best data scientist in the country, allowing them to perform optimally in utilizing data in order to find solutions for the banking sector,” explained Jahja Setiaatmadja, President Director of BCA.
The fifteen teams were shortlisted from an online competition conducted on the official website https://finhacks.id/. Starting on August 8, 2018, participants were challenged to develop machine learning models in three different categories, namely Credit Scoring, Fraud Detection, and ATM Cash Optimization. Through the Finhacks 2018 #DataChallenge, as many as 4.162 participants registered on the official website.
Not only that, the Finhacks 2018 #DataChallenge also organized three workshops in Jakarta, Yogyakarta, and Bandung, followed by 330 participants. The workshop invited keynote speakers from various technology companies in Indonesia, including Traveloka, Microsoft, DCI Indonesia, Algoritma, Tirto.id, Media Kernels Indonesia, and, of course, BCA.
Participants must take an online pre-assessment test, a mandatory to take part in the online data challenge stage. From here, 750 participants were shortlisted and grouped into 222 teams, generating 605 machine learning models submitted in the competition. After the stage ended on October 13, 2018, and passed the assessment stage by the organizer, 15 teams were selected to participate in the Demo Day in Jakarta and compete to win a total prize of Rp480 million.
In the Finhacks 2018 #DataChallenge Demo Day, the event was not only filled with presentations from the participants. There was also a panel discussion session called “Building a Great Data Culture”, attended by Armand Wahyudi Hartono, Deputy President Director of BCA, Crystal Widjaja, SVP Business Intelligence and Growth of GO-JEK Indonesia, and Tushar Bhatia, Head of Growth & Data Science of Bukalapak, with Rama Mamuaya, CEO of DailySocial.id as moderator.
The event was also enlivened by the performance of Dance of HaloBCA, Voice of HaloBCA, and special performance of stand-up comedy by Cak Lontong, who had brought laughter into the event at Soehanna Hall, Jakarta.
Disclosure: this article is an advertorial for the series of the Finhacks 2018 #DataChallenge supported by BCA.