Welcome back to the Agrifintech newsletter!
🤖 Traive AI credit breakthrough
Traive, the U.S. based and Brazil focused Agrifintech, announced a major AI model breakthrough. What did they achieve exactly and what might the implications be?
📣 It's all happening in Africa. Positive updates from AquaRech, Good Nature Agro, Khula, following on from Nile.ag in my last issue.
📣 AcreTrader broaden services and partner with FBN
Traive AI credit breakthrough
Traive is the first breed of a new generation of AI native fintech companies, that is rethinking the way the traditional fintech business model works.
What we have found is that the right AI system, with domain knowledge, can account for the challenges of data interoperability, so prevalent in Agriculture and other sectors. I believe verticals will be completely reshaped.
Aline Oliveira Pezente, co-founder of Traive.
Traive have just published two papers focused on the application of artificial intelligence (AI) to agricultural credit modelling and will present these to the ACM International Conference on AI in Finance this week (w/c 27th November) in New York.
I will let the AI community give their verdict on the impact on AI as I'm definitely not qualified to assess that. However, I did wonder about the impact on Agri Finance.
Last week, I briefly spoke with the team at Traive and carried into that conversation my (lazy Niall) thoughts on the topic, which mirrored those on blockchain - Proof of concepts are cute! Scaling data across value chains ... not so cute.
However, there were a few things that hooked me in. Firstly, the knowledge that the team have been working in this space for the past 5 years felt more like strategic product release than strategic press release for starters. Secondly, Traive had also developed the credit methodology behind one of my favourite Agrifintech structures, which I discuss below.
So let's look at
1️⃣ The key paper takeaways; and ask
2️⃣ So what?
Key paper takeaways
❶ Paper 1 - Agribusiness Delinquency Risk.
"Training a model [for automated credit risk assessment] requires immense data sets that are challenging to obtain" it read in the first few sentences. I'm simultaneously thinking 'here we go' 🙄 and curiously wondering how they've addressed this.
The research took datasets from 100,000 loans, sourced from 9 lenders and used this to train a Bayesian Network model for loan delinquency classification and found this offered up to 15% improvement on other data methods.
The Bayesian network model is key. In the financial context, lenders typically bypass data gaps by using 'experts' who grasp the context from experience. The Traive team used the Bayesian model as a proxy for the expert i.e. the technique models available data for relationships and infers there are correlations between latent, or missing variables.
🔑 Using data alone, this method offers an improvement of up to 15% and bridges the gap between the closed community of 'experts' in agricultural finance and the data teams. Scaling data is no longer a problem with this approach.
❷ Paper 2 - Enhancing Credit Risk Reports.
In Paper 1, guardrails are established for AI using the networked model, allowing it to deal with another major problem - hallucination (basically making stuff up). It's OK to hallucinate once in a while if you are using Chat GPT to suggest a complementary colour palette for your website or a recipe, but not for a lending decision with a few $$$ riding on it. 😅
According to the Traive VP for Data Science, Mohammed Ghassemi:
Paper 2 constrains the Chatbot so that it adheres to the modelled guardrails found in Paper 1 and prevents hallucinations downstream so that we get a higher fidelity model that can be used within a regulated environment.
Using the method of Labelled Guide Prompting (LGP), the Team fed the rules from the Bayesian network they established in Paper 1 to GPT-4 and found they were able to generate credit risk reports that were preferred in 60-90% of cases by human credit risk analysts.
🔑 We can feed the modelled Bayesian network guidelines to GPT-4 and generate credit reports which a human prefers. This demonstrates progress in solving the alignment problem of AI i.e. it can produce a result that a reasonable human would expect.
Ok... so what? Is there any impact on agricultural credit and a problem to solve?
To help steer me on this, I put out some thoughts to you guys on LinkedIn and people wanted to see use cases, funds deployed and, of course, a problem solved.
1️⃣ There are use Cases. As above, my first thought with AI in B2B credit was concern about data & use case scalability.
Fortunately, Traive had already soothed some of this anxiety through the development of the the credit methodology for the Responsible Commodities Facility an actual tangible, real world, use case - and one of my favourites.
This facility was established to help finance sustainably sourced soybeans from the Cerrado region of Brazil. The structure fascinated me as it pulled an interesting collective together - Farmers, SIM itself, end buyers in UK supermarkets (Waitrose, Sainsburys and Tesco no less), Collateral Managers and Traive to organise the data and make a recommendation to investor credit committees. See the structure below from the RCF annual report 👇
In its first year the facility raised $11m for farmers, structured this into a listed instrument on the Vienna Stock Exchange and finished the period with zero defaults. Last month, the structure increased in size to $47m and onboarded major banks including Santander and Rabobank to develop the structure.
Other Traive partners include FMC Corporation, who have used the models to develop credit structures for emerging farmers in Brazil.
Ok, this is not theoretical. In the overall scheme, $47m is still a small number but at least some credible names in the market trust it.
Am I hallucinating or does this actually work? You don't increase the size of a facility with a major question mark over some of the structure, do you?
2️⃣ Solving Problems in Financial value chains.
Where does this fit in exactly? What problems are solved?
🧐 The application of this to Capital Markets really jumps out. This is the downstream end of the financial value chain, after all loan documentation is in place and loans get pooled to be marketed to institutional investors. It is usually applicable to products that can aggregate standardised data e.g. credit card debt, mortgages.
The problem with doing this is data aggregation and standardisation. For example, Traive found up to 96% of some data points missing when they sourced data from lending institutions. How can anyone work with that? If the methods developed by Traive can avoid this, it allows for reduced transaction costs and increased liquidity in Ag related capital markets.
Previously, I quoted Fabricio Pezente from Traive:
In Brazil agribusiness is 30% of GDP and 5% of capital markets
I think developing the capital markets use case is key and this can definitely solve a problem. I want to refine that a little and ask the question - "where is this a problem?" Let's look at the U.S. and in turn other markets
🧐 Farmer Mac play a key role in this standardisation in the U.S., by buying approved loans from banks and 'drop shipping' these to capital markets investors.
Might they be interested in this? If they were to start from scratch tomorrow, yes. However, they have already invested in this onboarding process and building an origination network of hundred of Ag Banks.
But the main reason they may not be so enthusiastic is that their default rates are hideously low, with just $38m in cumulative loan losses since inception. A 15% improvement in this is a market opportunity of $5.7m. 😕 Still, there could be other problems to solve.
You could just build a digital native version of the existing process? Or work with the major cash rich banks who don't need to sell down loan books at all?
Many options are available and I look forward to see where Traive play in that particular market.
🤔 Unstructured Markets to benefit... Brazil and Mexico were cited by Traive as two markets where they are having the most success with financial actors - Lenders, Ag retailers and the large commodity buyers.
Not only are agriculture markets unstructured, but so too is the credit process and I think when we consider reduced transaction costs, reduced non performing loans and increased liqudity, this is where most impact will be felt.
Brazil is a gigantic market with lots of room for expansion and here the Traive team advised they work with "the 20% of producers who make up 80% of the production". Two of their use cases I have cited above are in this market.
Having worked in developing markets, similar to Mexico, I'm on the lookout for more progress here as things like logistics, physical infrastructure and government intervention start to play a role.
💭 ...But not unstructured value chains. There is a separate question around value chains.
In the Bayesian network model for example the Main Crop was the main determinant of the credit score, the market score and the agronomic score. The crops modelled were soy, summer corn, winter corn, wheat, rice, Arabic and robusta coffee.
It wasn't clear if this could be applied to different value chains such as fresh produce or livestock for example. I'm guessing the methods could be applied to alternative markets, but from a commercial perspective, maybe they just don't make sense yet.
3️⃣ An Infrastructure layer. The RCF above highlights the role for this type of product/ service in the overall financial stack, linking up with capital providers, data sources, products and distribution partners.
Traive have raised $30m+ from VC investors to date by developing and focusing solely on credit models. They are not building a bank, nor distribution, nor their own fund. There is plenty of room for others to do that and they need those counterparties to succeed.
It is clear from use cases above, partnerships are key. Will vertical finance get reshaped around these? I think it is going in that direction and now it feels like there is tooling available to nudge it along.
4️⃣ A Tenor issue. There was one obvious gap in the AI methodology presented - did you spot it?
The data studied and the applications so far are for short term working capital facilities. I've presented an outdated chart below from Farmer Mac but it raises an interesting point relevant to this conversation.
Real Estate related debt is 70% of the total debt with the remainder split between operating (short term working capital) and equipment.
The question remains whether the AI methodology developed can start to address the 70% for real estate and even the sizeable equipment finance market.
This obviously applies to U.S. and non U.S. markets (but we get good data from the U.S.)
It seems for now the main impact will be on operating credit.
- - - - - -
Now it's your turn - what do you think - Strategic Product or Strategic PR?
✅ Use cases in the market
✅ Capital deployed to short term working capital facilities
✅ Problems solved in the market (with others yet to tackle)
ℹ️ Data interoperability - need to test that one in practice.
Overall, I think this is very promising and cannot wait to see how markets develop in Brazil and Latam, but also in the U.S and other markets. As for other verticals, both within Ag (the livestocks etc) and outside of Ag, this also has to be promising.
- If you enjoyed this, do sign up - My next issue will be a shorter one covering 3-5 of the top questions I am asked. 😁 If you have questions, just shout, maybe it will make the list.
📣 It's all happening across Africa
⦿ Good Nature Agro, the Zambia based seed platform, which works with 30,000 growers to integrate them with the commercial seed market raised $8.5m to expand activities. Investment was led by Dutch based Goodwell Investments, who invested in an earlier round and they were joined by Oikocredit and Global Partnerships.
⦿ AquaRech, a Kenya based aquaculture platform received a $1.7m seed investment from Aqua-Spark with aprticipation from the Acumen Fund, Katapult and Mercy Corp Ventures.
⦿ Khula, a South Africa based digital agritech platform, receive investment from a PepsiCo South Africa backed fund. Terms or figures were not disclosed.
📣 AcreTrader expand services including partnership with FBN
The farmland investment sector is making the most of huge structural changes happening on the ground by introducing more entry and exit points into the asset class.
AcreTrader recently obtained their broker dealer licence from FINRA allowing them to structure more products and work with others in the sector. FBN, who have also established a business catering to farmland investment, have been announced as a new AT partner alongside Peoples Company and Strongwater Viticultural Investments.
The FBN connection has already allowed their investment business to support two option like structures for off market transactions, both of which were marketed via AcreTrader.
And that is everything this week folks - did you enjoy? I'm always here for some feedback.
If you enjoyed this please copy the link above and share on LinkedIn or with your internal network. 🙏