image: Figure 1: Logical flowchart of the MetaRL approach.
Credit: Sanjiv R Das
A team of researchers from Santa Clara University, the Indian Institute of Technology (IIT), and Franklin Templeton has introduced a breakthrough AI approach to Goals-Based Wealth Management (GBWM). Published in The Journal of Finance and Data Science, the study presents MetaRL, a meta reinforcement learning model that eliminates the need for separate, time-consuming optimization of the retirement plan for every individual investor.
"MetaRL provides near-optimal, individualized investment and goal-fulfillment strategies," says co-corresponding author Harshad Khadilkar, a visiting associate professor at IIT and director and principal research scientist at Franklin Templeton. "Pre-trained on thousands of goals-based wealth management scenarios, MetaRL achieves 97.8% of optimal utility while operating over 100 times faster than traditional methods, even under changing market regimes."
Traditional wealth management often relies on "one-size-fits-all" portfolios or on Monte Carlo simulation approaches that are very common but do not give truly optimal solutions. Dynamic programming (DP) models give truly optimal solutions but become computationally infeasible as financial problems become more complex. In contrast, MetaRL, inspired by the architecture of large language models (LLMs), is pre-trained on a large corpus of diverse financial planning scenarios. The system utilizes a unique environment-agent feedback loop to provide "zero-shot" solutions—providing immediate, optimized decisions regarding which goals to fulfill and how to adjust investment portfolios in real-time (Figure 1).
"Even when tested on capital market regimes it wasn't trained on, the model delivered expected utilities averaging about 98% of the theoretical optimum," shares Khadilkar. "Furthermore, MetaRL successfully addresses the "curse of dimensionality" that plagues methods like DP."
When incorporating stochastic inflation, DP requires at least four state variables, making optimization computationally infeasible. Whereas for MetaRL, its speed remained virtually unaffected even though it used 27 state variables.
"By eliminating separate training and optimization for each new investor problem, the MetaRL model produces near-optimal dynamic investment portfolios and goal-fulfilling strategies for a new problem within a few hundredths of a second," adds Khadilkar.
"This is just the beginning; this approach will be extended to more complex problems such as tax optimization as well," says co-corresponding author Daniel Ostrov from Santa Clara University. "This development marks a significant shift toward high-throughput, individualized financial planning, enabling institutions to offer sophisticated, dynamic strategies to a broader range of investors at unprecedented speeds."
###
Contact the author:
Sanjiv R Das, Professor of Finance and Data Science, Santa Clara University
Email: srdas@scu.edu
The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 200 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).
Journal
The Journal of Finance and Data Science
Method of Research
Computational simulation/modeling
Subject of Research
Not applicable
Article Title
A meta reinforcement learning approach to goals-based wealth management
COI Statement
We confirm that there are no conflicts of interest for any of the authors in this publication. There are no financial interests or personal relationships to be considered as potential competing interests. None of the authors was involved in the editorial review or the decision to publish this article.