Yapı Kredi Technology- Financial Virtual Assistant for Portfolio Management

FVAP

Financial Virtual Assistant for Portfolio Management

Yapı Kredi Teknoloji applied individually to the call titled "EuroHPC JU Benchmark Access".

Project Objective

Facilitating Access to Information on Mutual Funds

It is aimed to enable users visiting the Bank's website to access simple, clear and quick information on mutual funds.

Providing Personalized Portfolio Recommendations

Users will be supported in making more informed investment decisions by providing asset allocation recommendations in line with their risk profile and investment goals.

Providing Interactive Support with Virtual Assistant

A virtual assistant capable of providing quick and accurate answers to frequently asked questions about mutual funds and portfolio management will be developed.

Developing a Language Model Suitable for Turkish Investment Language

A special language model with high contextual integrity in Turkish will be created, capable of accurately interpreting terms and expressions specific to investment products.

Improving the Digital Investment Consultancy Experience

The investment consultancy service will be made scalable and accessible by making the investment process user-friendly in the digital environment.

Project Subject

It is aimed to develop an artificial intelligence-powered virtual assistant that provides users with information on mutual funds, possesses a command of Turkish investment terminology, and is suitable for both end-users and in-house use.

Providing Interactive Information on Mutual Funds

An intelligent virtual assistant capable of providing users with general fund information, historical returns, risk profiles and fund comparisons was developed.

Artificial Intelligence Model Compatible with Turkish Financial Language

Meaningful and accurate answers are provided to user's questions using a special language model trained with big data that comprehends Turkish investment terminology.

Multiplatform Compatibility

The virtual assistant was integrated into both areas accessible to end-users, such as the Bank's website, and in-house systems intended for employee use.

What Do We Do?

Developing the Artificial Intelligence Infrastructure

The leadership role was assumed in the architectural design and development of the project's artificial intelligence components.
Data processing and modeling strategies suitable for high-performance structures were developed.

Large Language Model (LLM) Development and Training

A large language model compatible with Turkish investment terminology was developed.
The LLM was trained on supercomputers to ensure high-accuracy output.

Transfer and Reinforcement Learning Applications

Transfer learning methods were applied to ensure the model's reusability across different institutions and data structures.
Reinforcement learning techniques were integrated to enable the model to evolve by learning from user interactions.

High-Performance Parallel Model Optimization

Parallel computing techniques were applied to ensure the model works efficiently in multi-GPU architectures.
Memory and processor optimizations were performed during the training and inference processes.

Project Output

Multiphase Language Models and Applications Optimized for Turkish Financial NLP

Language models specific to the financial sector, which improve the accuracy of financial texts and can be adapted to low-resource environments, were developed. Project outputs were documented with academic publications.

Multi-Stage NLP-Based Spelling and Grammar Correction System

The multi-stage NLP system used in financial reports and customer data in the field of portfolio management improved data quality by correcting spelling and grammar errors. This enabled more reliable analysis.

mT5-XL Model Optimization Providing Resource Efficiency

The mT5-XL model quantized with LoRA enabled real-time financial text analysis on low hardware resources. This optimization supported portfolio managers in making quick and economical decisions.

In-depth Interpretation of Financial Texts with FinancialBERT

The FinancialBERT model, trained with financial data sets, supported investment decisions by accurately interpreting market analyses and economic reports. The model provided accuracy in risk assessments with its high performance.