1. “The world needs banking,
but it does not need banks”
-Bill Gates
2. 2
Arboreum is an AI-
enabled platform for
receivables finance.
2
We harness the power of
“small-world networks”
within supply-chains to
unlock credit.
How?
1. We use supply-chain data to
underwrite risk
2. We use the supply-chain itself
to absorb part of the risk
3. 3
Hence networks are fundamental to our approach
We can holistically underwrite using
the structure of correlated time-series
on a network
• Decentralizable
• Respects privacy
• reliable credit risk assessment even if
data available for an individual borrower
is limited
We can distribute some of the risk onto
the supply-chain itself
• Goods on consignment represent existing
credit relationships.
• Owners of outstanding receivables can
earn yield simply by indicating their
acceptable default risk
• Securitized, diversified, senior tranche can
be sold ad-hoc to a market
4. 02
We enable formal credit to piggyback on informal credit
Tranche N
Tranche 2
Tranche 1
Arboreum investors
take lowest risk,
highly diversified
across loans
Anchor takes higher
risk at below-market
interest.
“Staked” excess
liquidity and NFTs
create the junior
tranche and
collateral
respectively.
Other businesses in the
value-chain can participate
in different tranches
Liquidity pool Investors purchase a
Collateralized Loan Obligation (CLO)
de-centrally minted and backed by
trade receivables.
4
100+ bn $USD in stablecoin currently earning no yield can safely earn 9-20% APR
financing real-world business activities
We are empowering businesses to finance themselves ad-hoc directly from
de-fi liquidity pools
Liquidity
Pool
NFT
NFT
NFT
1. A credit-worthy
“anchor” provides the
junior tranche
2. NFTs representing
the borrower’s
receivables are minted
3. The liquidity pool
provides remaining
required capital
Our AI and underwriting
tranches debt as to
maximize returns and
minimize risk for the
liquidity pool
5. Step 1: the credit union (in the market)
5
Arboreum uses the anchor’s
data to underwrite the
financing provided to union
members.
Credit union
Credit
union
Individuals and/or institutions
purchase senior–tranche,
securitized union debt.
Members borrow, preferably
against business they’re already
doing with the anchor (invoices,
salary slips, warehouse receipts).
A minimum of 20% if provided
through Junior-tranche debt which
comes from the union.
External
Lenders
Borrowers
Arboreum is building
credit unions as a
service, hinging on a
credit-worthy “anchor”
The credit union
serves as the junior-
tranche that co-lends
funds, and smooths
cash-flows by acting
as a ballast or buffer
of funds
The credit-union is
capitalized by
member-deposits and
Arboreum investors
6. Step 2: Tokenization of receivables (in progress)
6
Through their generous
support ($75K grant + $25K
credits) we are moving our
product onto the Algorand
blockchain.
ACH
1. Receivables can be
paid via multiple rails
(bank transfer,
credit/debit cards,
Automated Clearing
House).
Circle is merging the world of traditional banking with blockchains. Any
customer wishing to use our platform would be required to use Circle
for their receivables. Ergo:
2. Funds are collected in a Circle-
managed wallet denominated in
USDC.
3. Smart-contracts can be written
against the wallet, minting
various tokens on the Algorand
blockchain.
NFT
Receivables
Investment Asset
Reputation
4. Funds can be converted
to fiat either directly via
Circle or using any other
exchange
1. Receivables can be tokenized in a digital native manner.
2. We can accommodate far more complex structures than the credit-union.
We can organize credit along real-world supply chain networks.
7. Step 3: Full automation with decentralized AI
7
Just as networked AIs + Battery Storage have created the decentralized utility company;
we can create a true decentralized trade bank!
Each node-level AI needs just a few simple pieces of information:
1. Hurdle rate & target rate of return
2. Excess liquidity & receivables on hand
3. Max time period liquidity can be encumbered
4. Max acceptable discount rate on receivables
Once on a network the AIs in concert can solve a sophisticated
risk-management problem all through the process of novation:
1. Safely circuit excess liquidity and provisions collateral to
mitigate supply-chain frictions
2. Borrow funds/sell receivables directly from a market to
cover shortfalls
3. Enable a self-powered trade-finance system running on
blockchain + Arboreum Intelligence
We are building a plug-and-play solution to become the go-to
credit facilitator for independent networks such as marketplaces
We have built a
working Proof of
Concept of this
technology.
8. Our Current Core Assets
8
Receivables-financing front-
end and back-end application
(in production)
Conditional Copula Tail-
Expectation portfolio
optimization – proprietary
algorithm specifically
designed for the exigencies
of small, short-term loans
as the underlying basket of
assets.
Proprietary underwriting algorithms
based on the structure of correlated
time-series on a network.
“Tetris Algorithm” -
- easily tranches
payments and
mints CLOs with
almost any desired
number of tranches
and payment
schedule.
Algorithm is designed for
decentralization (please see
our whitepaper to learn more)
9. Gaurav Singhal | CEO
Lead Data Scientist, United Nations WFP
Machine Learning Engineer, Accenture
Technology Labs
Core Team
Junior Fellow, William Davidson Institute
Double MS, University of Michigan
(Economics & Information Science).
.
9
Dr. Avinash Bhardwaj | Technical Advisor
Professor of Engineering, IIT Bombay
PhD Operations Research, UC Berkeley
BTech, IIT Delhi
Dr. Igor Rivin | Technical Advisor
Former Regius Professor of Mathematics
Creator of Mathematica Kernel
https://en.wikipedia.org/wiki/Igor_
Rivin
PhD Mathematics, Princeton
Vishal Hemrajani | Engineering Lead
Engineering Lead, Valinor Earth
CTO, Flameback Capital
BE Electronics & Instrumentation, BITS Pilani
Julius Faber | Founding Engineer
Full-Stack Blockchain Dev, Neufund
Full-Stack Blockchain Dev, Consensys
MS Robotics & ML, TU Berlin
BS Computer Science, TU Berlin
Marc Held | Strategy Advisor
Built & sold multiple companies in the
Supply Chain + IOT space, currently
CEO & Founder, Fishtail.ai
BS Computer Science, Northeastern
11. Phase 1: Receivables Finance (Mostly Centralised)
The current goal is to build a scalable receivables finance solution for MSMEs connecting to
liquidity pools for funds
NFT
NFT
NFT
1. NFTs representing the
borrower’s receivables
are minted
Tranche N
Tranche 2
Tranche 1
NFT
NFT
NFT
NFT
NFT
NFT
2. Those NFTs get
tranched, pooled, &
divided to create
different CLOs
(centralized)
Liquidity pool
investors take lowest
risk, highly diversified
across loans
Anchor takes higher
risk at below-market
interest.
Other businesses in the
value-chain can participate
in different tranches
Liquidity
Pool
3. Different CLO tokens are minted with the
appropriate repayment terms and distribution of
profits controlled by smart-contract.
12. Phase 1: Fiat <-> Crypto is as its core
End borrowers get fiat and repay in fiat – goal is for them to operate their business insofar as
possible with minimal disruption. They issue invoices and expect to be paid in fiat.
Tranche N
Tranche 2
Tranche 1
Liquidity
Pool
1. When the loan is approved
stablecoin is automatically taken
from the liquidity pool
distributed in a manner as
specified by the original smart-
contract
2. Stablecoin is converted to fiat
and put into an escrow account
with a traditional bank.
3. Borrower receives
and repays in fiat.
4. Repaid funds is
converted back to
stablecoin.
5. Stablecoin + profits
distributed back to
tokenholders as originally
specified..
14. Arboreum holistically underwrites the borrower
Borrower
Supporter (degree 1 trustors) We let the network reach a “consensus” (not every
node must exactly agree) about the risk a borrower
poses over the tenor primarily by assessing cashflows
and correlation between cashflows.
15. Principles: Generalized Methods Always Win
The bitter lesson of 70 years of AI research: general methods leveraging computation are
the most effective, and by a large margin – incorporating human, domain-specific,
knowledge tends to over-complicate and under-perform human underwriting is obsolete
Principle 1: The best
predictor(s) (especially short-
term) of a given time-series
are other time-series’
Principle 2: In an ad-hoc
credit-network there exist
actually two kinds of risk we
must measure
1. Creditworthiness – what is
the probability that the
node returns the money
2. Credibility – how reliable
is the information a node is
providing on another node
Signal A Signal B Signal C
Beyond correlation, wavelets
allow us to study how one time-
series modulates another over
multiple different time scales
P(Credit-
worthine
ss)
P(Credi-
bility
P(Creditworthiness
| Credibility)
Principle 3: All knowledge is
Bayesian; everything is a
stochastic random variable
(i.e. its probability distributions
all the way down)
Specifically we use Beta
distributions to describe these
random-variables we are
building a Bayesian Reputation
System on an ad-hoc network
16. (1) Trusted Seed nodes incorporate off-chain data and
broadcast an initial estimate
Trusted nodes rate the borrower
and produce a rough estimate using
a reinforcement learning model
A
3
B
2
C
2
Borrower
Seed Node
We use a hazard-model in a reinforcement-
learning framework. We take a vector of
characteristics X, previous cash-flows Yt-1
and produce a best guess of loan default
over time for a given amount. X varies from
situation to situation. For example with our
logistics-partner pilot we have access to all
their historical consignment data.
17. (2) Each node makes its own prediction & computes
neighbour’s credibility Each node computes the creditworthiness
of the borrower using Bayesian wavelet
regression and forecasting
Each node computes its neighbour’s
credibility using a simple Bayesian model
relying on two inputs:
1. How close is my neighbour to the borrower?
This is determined from:
• Network distance between node and
borrower (the closer distance, the more
related they should be in real-life)
• Correlation of cash-flows with the borrower.
2. What is the historical performance of their
predictions?
Wavelets are a
powerful,
generalizable
time-series
modeling
technique that
work well with
minimal
assumptions
across different
temporal
scales. Wavelet
regression
methods often
out-perform
LSTMs while
matching their
black-box
convenience.
18. (3) The node updates the “weight” of its own forecast
(sensor fusion by Bayesian Kalman Filtering)
Each node is trying to
minimize its own
potential inaccuracy
compares the errors of
their neighbours’
models against its own
Forecast
Weight
19. (3) As the assessment propagates, each node updates
with its own assessment and broadcasts
Each node combines the creditworthiness message of their neighbours with
their own internal credibility measure of their neighbours
20. Better risk assessment
What is unique about
Arboreum’s risk assessment
Leverages the power of the network (the
signals provided by the credit-worthiness
of friends, the amount of trust that they
have placed, and the capacity of credit
flows on the network) to assess risk
Uses peer-enforcement mechanisms to
lower default risk by having friends
(trustors) participate in the loan by design
Continuously updates risk through
Machine Learning with Bayesian
modelling + Neural Networks; multiple
external and internal sources of
information used for this
Understands correlations between
loans/risk/borrowers e.g. what is the
probability you default if your friend's
friend defaults
~ 10% reduced risk of default
~ 80% increase in your default
probability
0.5 to 1.35% reduction in
interest rate for borrowers
5.5% relative increase in risk
adjusted returns for lenders
If a friend
contributes to
your loan…
It has been mathematically proven1 that large networks with
high connectivity result in lower interest for borrowers and
increased high-quality opportunities for lenders
If a friend
defaults...
Why bring in the societal trust and
network concept
1 literature can be shared upon request
~ 0.9 correlation in likelihood of
default if default risk was already low
22. Phase 1 & 2: Optimization-based Protocol
We currently implement a proprietary anti-fragile portfolio optimization approach
https://innolution.com/blog/agile-risk-management-antifragile
Goal: we want to create portfolios such that
volatility in the portfolio creates minimal
losses
https://goodup.com/who-will-turn-out-to-be-antifragile/
Antifragility is a characteristic of things
that gain from disorder
23. Antifragile Conditional Tail Dependence
One way to measure anti-fragility of a portfolio is the dispersion of the assets in the tails
https://analystprep.com/study-notes/frm/part-1/quantitative-analysis/correlations-and-copulas/
The portfolio with weak tail
dependence is more robust
than one with strong tail
dependence
This portfolio can be
considered anti-fragile
as dispersion increases
in the tails
We use a soft clustering
technique to determine weights
such that tail-dependence is
minimized
24. Phase 3: Neural Network-based Protocol
This essentially gives us the core functionality required to create not only truly decentralized credit-
systems but also general-purpose self-adaptive, intelligent networks that can solve distribution and
coordination problems.
In the future we will transition to a Reinforcement learning approach where portfolio weights
are dynamically learned from the joint-moments and spectral properties of the risk
distributions of assets
26. Risk distribution logic
Funds from node-to-node,
entities that know each
other off-chain and opt-in
to trust
This trust is the
fundamentally used as an
indicator of your
creditworthiness
Ergo: Default also affects
the credit worthiness of
those who opted to trust
you - a proven powerful
enforcement mechanism
Funds flow on the basis of
indicated trust
05
Lenders closest to the borrower:
• Have indicated trust in the
borrower
• Consequently, they bear the
highest risk and enjoy the best
interest rate from the loan
Lenders farther away from borrower:
• Don’t directly know the borrower
• So, their loan is more
collateralized so it can be safer
• Receive lower interest rates than
those close to borrower
Bob
i=8% c=20%
Eli
i=7% c=30%
Finn
i=4% c=60%
Numbers shown are interest rate i and
collateralization c as output of a simulation
27. What happens under the hood
At the heart of Arboreum is the ability for any blockchain node to (1) send messages, (2) coalesce messages; (3) compute new messages.
What makes the network intelligent? These messages can themselves be mathematical functions! Machine-learning algorithms compute these
messages from coalescing input messages and computing their desired optimal state, before passing the new message along.
AI
Message
Function Description
1
Assess
Risk
2
Collate &
Compute
Demand
Function
3
Novation
Alice
Bob
Charlie
Eric
Donna
The respective
borrowing/lending
histories of nodes
and the trust
amounts are used
Wave 3
to construct a probability distribution
representing their best guess of the risk
of default.
compute a function equating the
amount of the loan they will buy
given rate (R), tenor (T) and
collateralization level (C).
Given a view of the
loan risk and their
neighbors' demand
functions as well,
the AI can then
i.e. as the loan is handed off in smaller
and smaller amounts from neighbor to
neighbor the rate, tenor, and collateral
are altered to decrease risk in
exchange for less interest.
R1
C1
T1
R2
C2
T2
The bond is
propagated
through the
process of
novation;
Wants A
$1000 Loan
Commits $200
@ r=8%,c=20%
Commits $50
@ r=5%,c=40%
Receives Loan
@ r=8%,c=20%
Eric decides to
not participate
due to risky terms
Commits $25
@ r=4%,c=60%
These messages propagate across the network in 3 consecutive
waves, allowing the network to intelligently distribute the loan
Wave 1
Risk Assessment
Messages
Bob assesses Alice's risk profile and passes that as a
message to Charlie, who repeats this whilst adding in
the information he knows about Bob. This continues
up the chain until the message reaches Eric.
Wave 2
Demand Function
Messages
Given his view of Alice's risk, Eric computes his
demand function and passes it along as a message to
Donna who does the same, knowing there is a
possibility to novate some of Alice's loan to Eric. This
continues back up the chain until it reaches Alice.
Wave 3
Novation
Messages
Alice chooses from the options presented to her by
combining the demand functions of her neighbors,
then the loan is novated back up the chain and
accounts are adjusted.
Wave 1
Wave 2
28. Under the Hood – An Example Loan
06
M
A6
A5
A4
A3
A2
A1
B6
B1 B2
B3
B4
B5
C1
C2
C3
C4
C5
C6
$200,
r=8%,c=20%
r represents the
interest rate, and c
represents the
collateralization ratio
$100,
r=8%,c=20%
$50,
r=4%,c=60%
As seen in this example, when the loan
propagates it is novated as follows:
A chain of accountability from creditor to borrower is created,
resulting in a proven, highly-effective, peer-enforcement
mechanism that incentivizes loan repayment.
Interest Rate
Collateralization
Rate
Risk-Adjusted Return
Steps Away From Loan Originator
Hence when default occurs, losses to
lenders accrue as follows:
Collateral Amount
% Of Loan Held By
Lender (Principal Owed)
Losses
Steps Away From Loan Originator
29. Less risk and lower spreads
Our simulated network
consists of 125 people, each
with different risk tolerances
and amount in wallet
In all scenarios, we assume
that the borrower is Ram, he
wants to borrow INR 25K,
for a duration of 1 year, is
willing to pay an interest
rate of at max 20%, and he
has 5 immediate trustors
on the network
We make the assumption
that Ram is comfortable
committing 25% as
guarantee in case he
defaults1
1
Social lending scenarios varied by degrees of separation from borrower
4 5
2 3
Number of loan
contributors
Interest rate, %
Average
individual risk
of immediate
trustors, INR
For Ram
For network
Simulation outputs for a single loan
1 Details of how to manage guarantee commitment while respecting RBI regulation to discussed
40%
22%
14% 12%
5 20 27
42
110
4000 3300
2250 2000
31
17 13 11
Risk-adjusted
return, %
-9
-5
-1 -1
AI solves for optimal risk-adjusted distribution of any loan over the network, while respecting the individual risk tolerances of all in the
network. Benefits increase as the loan spreads farther and farther.
Risk premium as
adjustment to
interest rate, %
n/a as
only INR
10k are
available
from
Ram’s
immediate
trustors
n/a
AI will propose 12% interest rate as it meets Ram’s
requirement while giving all lenders significantly safer loans
This is the power of social lending. Smart contracts will
spread the loan over people till 5 deg of separation away.