Lalit, Ed and AI/M team - Nice job on: https://mifosforge.jira.com/l/c/8KpdQP4a
This is a nice compilation of potential items for AI/ML on the Fineract/Mifos stack. #1) I think it would be great to turn this into a White Paper for the fineract/mifos communities. The White Paper should also address where in the stack changes are needed. I believe that "data pools" (or similar) would need to be created outside of the operational datasets, and that may require some changes to the database and data extraction strategies. I guess there are many issues that need to be addressed and reasons to move these functionalities forward. Making that case formally and determining criteria for priorities, seems like a good step. #2) In the absence of an overarching framework for evaluating priorities, my gut instinct on priority: A) help with operational risk (e.g. Fraud, Portfolio risk factors, projection of on lending or capital requirements ) ; B) improve product reach (e.g. more Credit Risk scoring); C) make operations more efficient . (in that order) #3) I would add one Use Case, incorporating into the money management use case the concept of multi-currency and market fluctuations for reducing exposures. There are two applications of this. *ONE* is that many Financial Institutions (including Microfinance Orgs) take out debt for on-lending in dollar or euro accounts and have to contend with repayment in local currencies, thus requiring careful tuning of their interest rates charges to consumers and other currency hedges. *TWO* some financial institutions are participating in remittance schemes where FX exposures are non-negligible, and some FIs would anticipate being in an intermediary role in those flows if they could. #4) Datasets - I only have some suggestions - areas of inquiry: A) better internal data: part of the issue with fraud detection is finding the right sort of pattern recognition - and that requires looking at a lot of operational data (timing of loans, amounts, unusual transfers, login from devices, etc) and then flagging potential cases for human review. Algorithms can then be trained. B) economics for products and risk of portfolio require exploring available proxy data. The "people's economy" - i.e. the economy lived by the poor or semi-poor often is obscured from official statistics. Call Data Records (CDR) were an early area of exploration but for obvious reasons the mobile networks are not keen to share that. Consumer spending data for things like kerosene, wood for cookfires, LPG, motorcycles, bikes, solar lanterns, may be a good way to go if available. Commodity prices are useful in anticipating consumer spending reductions in other areas (i.e. the price of rice goes up, spending for other consumables goes down ... in theory) I hope all that helps. Thanks, @[email protected] <[email protected]> On Wed, Jan 20, 2021 at 10:26 PM Ed Cable <[email protected]> wrote: > Hi everyone, lalit and the other members of our AI and ML working group > have been documenting the various AI use cases that we could possibly focus > on as part of the community-wide AI for all strategy and roadmap. > > We would like to get the community's feedback on these use cases and which > ones you might like to see prioritized, and whether or not you've got data > sets to help with some of the use cases that are focused upon. > > Could you please review https://mifosforge.jira.com/l/c/8KpdQP4a? > > > And share your comments via email on the following (especially the > priority of which use cases to focus on): > > a) Addition/modification of use cases > b) Priority of the use cases > c) Available datasets and any other inputs on federated learning > > > The working group meets every other Friday for those interested in > participating. > > Best wishes, > > Ed > > > >
