Haute Lumière
Commerce · III.06 · MMXXVI · daylight
Mycelial Finance
Volume III — Money, Energy, Information
Nine movements, one intermediary.
THE PLATE
THE LETTER
There is a version of this chapter that would be very easy to write and would be worthless. It goes: the forest is a network, the network shares, therefore finance should share, therefore here is a beautiful diagram. It has been written many times. It is the reason a serious reader puts a book like this down.
So we are going to do something slower and much more useful. We are going to find out what the biology actually says in 2026 — including the parts that have been retracted in substance by the people best placed to know — and then build the finance only on the parts that hold. You will end up with fewer metaphors and one instrument you could take to a credit committee.
The reason this is worth the trouble is that the true version of the mycorrhizal story is a better guide to finance than the popular one. The popular version is about generosity. The true version is about a market with an intermediary that takes a position, discriminates between counterparties, holds inventory, and keeps a spread. That is not a disappointment for anybody building a regenerative financial instrument. It is the design.
You already know why this matters commercially. Almost every attempt to move capital in small amounts, to many people, on local information, has run into the same wall: the cost of deciding yes or no does not fall with the size of the loan. A network can route money beautifully and still die of the cost of refusing somebody. This chapter prices that wall exactly, and then shows the three structures that get under it.
One honest warning about scope. Nothing here proves that a distributed lender outperforms a bank. What the arithmetic supports is narrower and more useful: there is a range of transaction sizes where centralised assessment cannot operate at all, and in that range a network with local signal is not the romantic option — it is the only option, and it has a price.
— The Editors
DISCOVERY
What is already working
The biology, at its best.
In 2016 Christian Körner's group at Basel did something that had not been done before: they labelled the crowns of five 40-metre Norway spruce with tagged carbon dioxide from a construction crane and followed it for five years. The label turned up in the fine roots of neighbouring beech, larch and pine. Roughly 40 percent of fine-root carbon in those trees came from a neighbour, and the flow ran in both directions at an estimated 280 kilograms of carbon per hectare per year.
That is a measurement, taken in a real forest, on mature trees, over five years. Take it seriously. Note also the word the authors themselves chose for the title: trade.
Alongside it sits the most useful mycorrhizal result of the century for anybody who thinks about capital. Toby Kiers and colleagues showed in 2011 that the plant and the fungus reward each other in proportion to delivery: a root sends more carbon to the fungal partner supplying more phosphorus, and the fungus sends more phosphorus to the root supplying more carbon. Neither is generous. Both discriminate. In 2019 Whiteside and colleagues went further and tagged phosphorus with quantum dots: the fungus moves phosphorus from where it is abundant to where it is scarce, and trades it there at a better rate. That is arbitrage, performed by an organism without a nervous system, and it is beautiful.
The finance, at its best.
Four cases where the same shape already earns money.
Zopa, London, 2005–2021. The first peer-to-peer lender. Its most remarkable period is the one nobody markets: through the 2008 financial crisis its lenders kept positive net returns while the banks around it were being recapitalised by the state. Small amounts, many borrowers, no branch network.
Kiva. Loans assembled in $25 pieces and disbursed through local field partners who do the assessment on the ground. The published historic repayment rate is around 96 percent. The field partner is not a nicety in that structure; it is the structure, and we will price it in the next movement.
Revenue-based financing. A live, priced, institutional market. Lighter Capital publishes its terms: 2 to 8 percent of monthly revenue until a 1.35× to 2.0× repayment cap is reached, on facilities up to $4,000,000. No board seat, no equity, no personal guarantee — and, crucially, no covenant tested against a central assessment of the borrower's quality. The payment is indexed to the borrower's own local signal.
The rotating savings association. Documented by Shirley Ardener in 1964 and older than any of the above: a susu, a tanda, a chit, a hui. Every member pays in each period; one member takes the pot; the cycle continues until everyone has taken once. Zero interest, zero central underwriting, and default rates that embarrass the formal sector — because the assessment is done by people who will see you on Saturday.
Four instruments on four continents, one pattern: each of them pushed the decision to the edge and kept the accounting in the middle. That is the mycelial move, and it is not a metaphor — it is a topology, and topologies have costs.
THE ARITHMETIC
What survives, what does not, and what it costs to run
First: what the biology will actually support in 2026.
In February 2023 Justine Karst, Melanie Jones and Jason Hoeksema published a review in Nature Ecology & Evolution that examined three claims made constantly in popular writing about the "wood wide web". It is the most important paper in this chapter and it is inconvenient in exactly the right way. Here is what it found, claim by claim.
Claim one — common mycorrhizal networks are widespread in forests. Insufficiently supported. Genotype mapping of network structure is arduous, and it has been done in only five studies, across two forest types.
tree species mapped in relation to a CMN 2
tree species on Earth (Cazzolla Gatti et al., 2022) 73,300
-----------------------------------------------------------
share of the world's trees mapped 0.00273 %Three ectomycorrhizal fungal species have been mapped in relation to a network, and two studies demonstrate actual continuity of fungal links between trees. Shared species on adjacent roots is not a shared individual, and only a shared individual is a connection.
Claim two — resources move through those networks and improve seedling performance. Insufficiently supported. The review evaluated 26 field studies. Seventy-seven percent were in ectomycorrhizal forests, and 35 percent of those were on Douglas fir:
field studies reviewed 26
in ectomycorrhizal forests 26 x 77% = 20
of those, on Douglas fir 20 x 35% = 7Most ran two years or less; one ran five. And for every study in that set, the authors concluded, the result can be explained without invoking a network at all — because a solid barrier that blocks hyphae also blocks soil solution, and soil solution carries resources perfectly well.
Claim three — mature trees preferentially send resources and defence signals to their own offspring. No peer-reviewed published evidence. Not weak evidence. None. The single peer-reviewed greenhouse study testing kin effects found the labelled carbon travelling through the soil solution rather than through a network.
Then the authors did something braver and audited how this literature is cited.
papers citing 7 influential studies on network structure 593
papers citing 11 influential studies on network function 1,083
-----------------------------------------------------------------
papers audited 1,676
unsupported citation rate, structure, by the end of period ~25 %
unsupported citation rate, function, by the end of period ~50 %
growth in unsupported citations, structure 1.047 x per year
compounded over 25 years 3.15 xUnsupported citation of this literature roughly tripled in twenty-five years. If the terminal rates held across the whole audited set — an upper bound, not an average, because the rate rose over time — it would touch 690 papers, 41 percent of the literature.
So, precisely: carbon and nutrients demonstrably move between plants through soil, and some of it moves through fungi. What does not survive is the direction of intent. The tree is not giving. The fungus is not delivering. And the mature tree feeding its children is, at the time of writing, a story.
Put Klein's measurement next to that and the honest reading is clear. Against a temperate forest net primary production of 7,000 to 12,000 kg C per hectare per year, 280 kg is:
280 / 12,000 = 2.33 % 280 / 10,000 = 2.80 %
280 / 7,000 = 4.00 %Two to four percent of the forest's annual carbon, moving sideways, in both directions, by a route nobody has yet isolated. Real. Bilateral. Small.
Second: the number that governs every distributed lender ever built.
The cost of deciding whether to lend does not scale with the loan. A relationship officer costs roughly the same to assess $20,000 as $2,000,000 — which is why, as Karen Mills and Brayden McCarthy documented, banks quietly abandoned small-business lending below a few hundred thousand dollars. Put a figure on it and the consequence is arithmetic, not opinion.
assessment cost, relationship officer $3,000 (assumed; range 2,500-4,000)
assessment cost, decision engine $25 (assumed)
assessment cost, a neighbour: 2 h x $20 $40 (assumed, and IN KIND)
net spread kept 6.0 % per year
average duration 1.5 years
margin per unit of principal 9.0 % = 6.0% x 1.5Breakeven ticket is assessment cost divided by margin:
officer $3,000 / 0.09 = $33,333
neighbour $40 / 0.09 = $444
machine $25 / 0.09 = $278A person cannot assess below about thirty-three thousand dollars. A machine reaches two hundred and seventy-eight. Every network in the Discovery movement is an answer to that one line. Kiva's field partner, Grameen's group of five, the susu meeting: all of them are ways of buying assessment at $40 instead of $3,000.
Third: what that $40 actually costs, priced honestly.
A joint-liability group of five meeting weekly for an hour:
person-hours per group per year 5 x 52 x 1 = 260 hours
group book 5 x $300 = $1,500
at $1.00/hour shadow wage 260 x 1.00 = $260/yr
$260 / $1,500 = 17.3 % of principalGrameen's stated basic rate is 20 percent on a declining balance. The unpriced labour in the method is of the same order as the interest charged. At $0.50 an hour it is 8.7 percent; at $2.00 an hour, 34.7 percent. Local assessment is not cheap. It is unbilled, and the bill is paid in the borrowers' Tuesday evenings.
Fourth — the honest negative, and it is the load-bearing one.
A network with no central assessment has no way to refuse a bad borrower cheaply. Say the local signal is good: it catches three-quarters of bad borrowers and passes 85 percent of good ones, out of a pool with an 18 percent bad rate.
1,000 applicants bad 180 good 820
bad, passed anyway 180 x 25% = 45
good, passed 820 x 85% = 697 approved = 742
good, REFUSED 820 x 15% = 123
post-screen bad rate 45 / 742 = 6.06 %The signal works: expected loss falls from 11.7 percent of principal to 3.94 percent — 7.76 points. And the price of it is 123 good borrowers turned away, 0.91 good refused for every bad caught, which on a $1,484,000 book is 1.49 points of forgone margin. Net, the signal is worth +6.27 points, and that is why these structures exist at all.
But look at what refusal is in a network. A bureau refuses 1,000 applicants for $25,000 — 1.68 percent of the book — and spends nothing else. A node refuses by telling a neighbour no. There is no invoice for that and there is no second one available. Which sets the ceiling:
stable relationships per person (Dunbar, 1992) 150
plausible borrowers among them, at 30% 45
book per node 45 x $2,000 $90,000
gross margin $90,000 x 6% $5,400 / yr
servicing 45 x $40 $1,800 / yr
net to the node $3,600 / yrThree thousand six hundred dollars a year. That is not a job. A node cannot live on the spread its relationships can carry, which is why every network of this shape either pays its nodes in something other than money — status, reciprocity, access, membership — or hands the economics to an intermediary large enough to own the software. There is no third case in the record.
Fifth: redundancy, priced — where the forest analogy actually fails.
The living-systems argument says redundancy beats efficiency: many small positions survive what one large one does not. It is true, and it is true only against idiosyncratic loss. Take 10,000 loans, an 8 percent default rate and a 65 percent loss given default, and vary only the correlation between borrowers:
expected loss 8% x 65% = 5.2 %
one loan's loss sd 0.65 x sqrt(0.08 x 0.92) = 17.63 %
portfolio sd = 17.63% x sqrt( 1/N + rho(1 - 1/N) )
rho = 0.00 sd 0.176 % 1-in-100 loss 5.61 %
rho = 0.02 sd 2.500 % 1-in-100 loss 11.01 %
rho = 0.05 sd 3.947 % 1-in-100 loss 14.38 %
rho = 0.15 sd 6.832 % 1-in-100 loss 21.09 %Correlation of 0.05 multiplies the volatility by 22 times. And here is the sentence to carry: at that correlation, going from 1,000 borrowers to 10,000 — ten times the diversification, ten times the origination cost — reduces the standard deviation by 0.03 percentage points. It buys nothing.
That is the threshold below which this whole approach fails, stated exactly. A distributed network is a superb defence against one borrower failing and a negligible defence against a drought, a rate shock, or a regional downturn. The forest has precisely the same failure mode: a stand absorbs a dead tree and does not absorb the fire. RateSetter's provision fund — a genuinely well-designed mutual reserve — met that wall in May 2020, and the response was to halve lender interest rates for the rest of the year and divert half of all returns into the fund. Redundancy is never free. It is paid for out of somebody's coupon, and the design question is only ever whose.
DREAM
What becomes ordinary
Describe it in the present tense, because a dream in the future tense is a wish.
In the economy that has learned this, capital reaches small enterprises without first being converted into a credit score. A supplier network of four hundred growers has a facility inside it: each grower's repayment is a fixed share of what they actually sell, the cooperative holds the ledger, and nobody is asked to produce three years of audited accounts to borrow eleven thousand dollars for a pump. The assessment is done by the buyer who has taken their crop for six seasons, and the cost of that assessment is a phone call, because the buyer was going to make the call anyway.
Intermediaries take positions. This is the part that looks strange from here and is entirely ordinary there. The platform in the middle of a lending network does not present itself as a neutral pipe and then disappear when losses arrive; it holds a vertical slice of every facility it originates, on its own balance sheet, and it is paid a spread rather than a fee. Everyone can see what it keeps. Everyone can see what it loses. The disclosure is one line and the alignment is complete.
Correlation is a disclosed number. Every facility that pools small exposures publishes the correlation it has assumed, the evidence for that figure, and what the book does if the figure is wrong by a factor of two. It sits on the front page next to the expected loss, in the same size type, because it is the assumption that actually decides the outcome. Nobody finds this remarkable. It is simply how these things are written.
And the small transfer is unremarkable. Moving $400 to somebody eight hundred miles away, against a claim on their future revenue, costs a few cents and settles in a second. The floor that used to sit at thirty-three thousand dollars sits below five hundred, which means a whole layer of economic life — the repair shop, the seed-cleaning cooperative, the two-person studio, the person with a van — is inside the financial system rather than beside it.
None of that requires a change in the law, a new technology, or anybody's conversion. It requires an instrument whose assessment cost is under forty dollars, an intermediary willing to be on risk, and a correlation figure written down where a reader can argue with it.
DESIGN
The structure that gets there
Four components. They are not optional and they are not independent.
One — the local signal, and a definition of it that a machine can read. A signal is only usable if it arrives without a person having to fetch it. Revenue through a payment processor, volume through a cooperative's own weighbridge, invoices through a supply-chain portal, energy exported to a grid. The test is brutal and simple: can the lender see it without asking? A signal that requires a request has an assessment cost, and we have already computed what an assessment cost does. This is why revenue-based financing worked commercially and why so many "community lending" pilots did not.
Two — payment indexed to the signal, never to a covenant. The borrower pays a share of what moves, not a fixed instalment against a forecast. This is the single most valuable thing the biology suggests and it survives every critique in the Arithmetic: in Kiers' experiments the exchange rate follows delivery, period by period, with no assessment of intent at all. The consequence in finance is that a bad year is paid for in time rather than in default, which is worth computing precisely:
advance $250,000, cap 1.5x = $375,000, share 6% of revenue
revenue starts at $150,000/month, so payment one is $9,000
revenue +2%/month repaid in 31 months (2.6 yr) IRR 34.6 %
revenue flat repaid in 42 months (3.5 yr) IRR 27.7 %
revenue -1%/month repaid in 54 months (4.5 yr) IRR 23.7 %The borrower shrinks by a quarter over four years and still repays in full. The lender's loss is yield, not principal — 34.6 percent down to 23.7 percent. No workout, no restructuring, no covenant breach, no lawyer.
Three — a first-loss reserve, funded from origination, sized against correlation and not against the mean. Mutualised reserves work. RateSetter's did, for a decade. What breaks them is sizing to expected loss when the loss that matters is the correlated one. Fund it out of the origination fee so it accumulates with the book rather than requiring a decision, publish its coverage ratio monthly, and state what it does not cover in the same sentence.
Four — an intermediary that holds a vertical slice. This is the component the first generation of platforms refused, and it is the one the whole record insists on. A pipe that earns a fee on volume and bears none of the loss will originate volume. Require the intermediary to retain a slice of every facility, pari passu, on its own balance sheet — and then let it earn a spread rather than apologise for one.
Sequence. Signal first, because everything else is priced off it. Then the indexed payment, which can be tested on ten facilities with a spreadsheet. Then the reserve, once there is a book to reserve against. The intermediary's retention is written into the first document and never retrofitted, because a platform that has been fee-earning for two years cannot be asked to start holding risk.
DESTINY
How it holds when nobody is pushing
It holds when the intermediary cannot separate its own outcome from the book's. That is the whole of it, and the record is unambiguous.
Between 2020 and 2021 the three most serious peer-to-peer lenders in the English -speaking world all resolved the same way. RateSetter sold its loan book to Metro Bank in September 2020 and closed investor accounts on 2 April 2021. LendingClub bought Radius Bank and retired its retail Notes platform on 31 December 2020. Zopa, the original, closed every investor account on 7 December 2021 and became a bank.
The usual reading is that peer-to-peer failed. It is the wrong reading, and correcting it is the point of this chapter.
None of them abandoned distributed origination. All three kept it. What they abandoned was the pretence that the thing in the middle was a pipe. They became principals: taking deposits, holding the loans, bearing the loss, earning the spread. And a balance-sheet lender making thousands of small loans on automated signal is the mycorrhizal structure — a discriminating intermediary, on risk, running its own book between many small counterparties.
Here is where it fails. It fails when the intermediary earns a fee rather than a spread, because then volume is the only variable it controls. It fails when the reserve is sized to the mean and the loss arrives correlated. It fails when the local signal quietly becomes a form somebody fills in, which converts a $40 assessment back into a $3,000 one without anybody noticing, and the network dies of a cost nobody put in the model. And it fails when the nodes are asked to work for the spread their 150 relationships can carry, which we computed at $3,600 a year — so if a network's plan requires its nodes to be paid in money, the plan is already wrong and the arithmetic said so before the first loan.
DELIGHT
What it feels like
The pleasure here is the pleasure of a thing turning out to be more interesting once the decoration comes off. You arrive believing the forest is generous. You leave knowing it is a market — with arbitrage, inventory, counterparty discrimination and a spread — and the market turns out to be four hundred million years old and still running, which is a far better recommendation than kindness would have been.
There is a second pleasure, quieter, in watching a repayment arrive that is smaller than last month's because the borrower had a thin month, and nothing whatever happening. No letter. No default notice. No flinch. The instrument absorbed it because it was built to. Somebody's difficult August passes through a financial structure and comes out the other side as a slightly longer term, and that is the entire event.
And then the specific joy of a hyphal fan under a lifted root: too fine to photograph well, doing arithmetic that nobody taught it, entirely without sentiment, and keeping the whole stand alive anyway.
OPERATIONALIZE THIS
At the level of finance
The instrument: the Hyphal Facility — a revenue-indexed micro-facility, originated through local nodes, with a mutualised first-loss tranche and a sponsor that retains a vertical slice.
The structure. A sponsor originates facilities of $2,000 to $50,000 through nodes — a cooperative, a trade association, a franchisor, an anchor buyer — who supply the local signal and receive an origination share. Repayment is a fixed percentage of the borrower's observed revenue until a cap of 1.35× to 1.5× is met. No covenant, no personal guarantee, no fixed maturity. A first-loss fund of 6 percent of principal is funded out of origination. The sponsor retains 10 percent of every facility, pari passu.
The balance-sheet treatment, and this is where a treasurer will stop you. Under IFRS 9 a financial asset is held at amortised cost only if its cash flows are solely payments of principal and interest. A payment indexed to revenue is not, so these assets sit at fair value through profit or loss, and the P&L will move with the revaluation. Do not discover this in the audit. Model it in month one, agree the valuation methodology with your auditors before the first drawdown, and present the volatility as what it is — the price of an instrument that cannot default the way a loan defaults. Under US GAAP the equivalent conversation is ASC 825. Have it early and it is a paragraph; have it late and it is a restatement.
The counterparty. Your own supply chain first, always. You already have the signal — you see their invoices — and you already have the relationship, so the assessment cost is genuinely near zero rather than theoretically near zero. A $1,200,000 supplier facility at a $25,000 average ticket is 48 counterparties: enough to be a portfolio, small enough to run on a spreadsheet. At a 6 percent net spread that book earns $72,000 a year, and at the 5.2 percent expected loss it absorbs $62,400 — margins this thin are exactly why the correlation figure is the decision, not a footnote.
The number that decides it. Not the yield. The assumed default correlation, ρ.
gross yield 22.0 %
expected loss (rho = 0) 5.2 %
servicing 3.5 %
origination, amortised 2.5 %
------------------------------------------------
net 10.8 % vs cost of funds 7.0 %
margin +3.8 points
the same book, one year in a hundred, at rho = 0.05:
loss 14.38 % -> net 1.62 % vs 7.0 % shortfall -5.38 pointsOne bad year costs 1.42 good years of margin. Unexpected loss is 9.18 points, the fund covers 6.0, and the remaining 3.18 percent of the book is equity the sponsor must hold. Put ρ on the front page with the evidence for the figure you chose. If you cannot source it, say so and use 0.10; a stated assumption that is too conservative costs you a few points of return, and an unstated one costs the fund.
The first ninety days.
| Day | Action | Artifact |
|---|---|---|
| 1–15 | Name the signal and prove you can read it without asking | A data feed, running |
| 16–30 | Price the assessment cost per facility, honestly, including in-kind time | The cost-per-decision line |
| 31–45 | Agree the ρ you will underwrite to, and its evidence | The correlation memo |
| 46–60 | Settle IFRS 9 / ASC 825 classification with the auditor | A signed accounting note |
| 61–75 | Ten facilities, one node, real money, 10 percent retained | Ten signed agreements |
| 76–90 | First repayment cycle; publish the fund's coverage ratio | The monthly one-pager |
Ten facilities is the right size. It is above noise, below any committee's patience for a second meeting, and it produces the only thing that matters next: a repayment series that somebody else can check.
APPRECIATIVE QUESTIONS
Twelve, for a room
Discovery — what is already working
- Where in this organisation do we already lend — in cash, in materials, in terms, in time — to somebody we would never formally underwrite? Who decided that, and what did they know?
- Which of our counterparties do we see so clearly that we could predict their next quarter without asking them? What are we seeing, exactly?
- When has extending terms to someone in a thin month come back to us later? Tell the specific story, and say what it was worth.
Dream — what becomes possible
- If the smallest amount we could sensibly advance fell from thirty thousand to five hundred, who would we be doing business with next year that we are not doing business with now?
- Imagine our credit paper with the correlation assumption on the front page in the same size type as the yield. What conversation does that room have that it does not have today?
- If every repayment we received moved with our counterparty's actual week, what would our collections function do instead of collecting?
Design — what we build
- What signal about our suppliers do we already receive, without asking, that we currently throw away?
- Which of our relationships could carry an assessment at forty dollars — and who exactly would be doing it, and what would they want in return?
- What is the smallest vertical slice we could hold in something we originate, that would change how carefully we originated it?
Destiny — how it holds
- What would have to be true for this facility to still be running when nobody who designed it is here?
- If our correlation assumption turned out to be wrong by a factor of two, who finds out first, and by what route?
- What would we see in the first month of this quietly turning back into a form somebody fills in — and who is closest to noticing it?
WORKS CITED
Karst, J., Jones, M. D. and Hoeksema, J. D. (2023). "Positive citation bias and overinterpreted results lead to misinformation on common mycorrhizal networks in forests." Nature Ecology & Evolution, 7, 501–511.
Simard, S. W., Perry, D. A., Jones, M. D., Myrold, D. D., Durall, D. M. and Molina, R. (1997). "Net transfer of carbon between ectomycorrhizal tree species in the field." Nature, 388, 579–582.
Read, D. J. (1997). "The ties that bind." Nature, 388, 517–518.
Klein, T., Siegwolf, R. T. W. and Körner, C. (2016). "Belowground carbon trade among tall trees in a temperate forest." Science, 352(6283), 342–344.
Kiers, E. T. et al. (2011). "Reciprocal rewards stabilize cooperation in the mycorrhizal symbiosis." Science, 333(6044), 880–882.
Whiteside, M. D. et al. (2019). "Mycorrhizal fungi respond to resource inequality by moving phosphorus from rich to poor patches across networks." Current Biology, 29(12), 2043–2050.
Robinson, D. and Fitter, A. (1999). "The magnitudes and control of carbon transfer between plants linked by a common mycorrhizal network." Journal of Experimental Botany, 50(330), 9–13.
Henriksson, N., Marshall, J., Högberg, M. N. et al. (2023). "Re-examining the evidence for the mother tree hypothesis — resource sharing among trees via ectomycorrhizal networks." New Phytologist, 239(1), 19–28.
Cazzolla Gatti, R. et al. (2022). "The number of tree species on Earth." Proceedings of the National Academy of Sciences, 119(6), e2115329119.
Simard, S. (2021). Finding the Mother Tree. Knopf.
Ardener, S. (1964). "The Comparative Study of Rotating Credit Associations." Journal of the Royal Anthropological Institute, 94(2), 201–229.
Stiglitz, J. E. and Weiss, A. (1981). "Credit Rationing in Markets with Imperfect Information." American Economic Review, 71(3), 393–410.
Akerlof, G. A. (1970). "The Market for 'Lemons'." Quarterly Journal of Economics, 84(3), 488–500.
Ghatak, M. and Guinnane, T. W. (1999). "The economics of lending with joint liability." Journal of Development Economics, 60(1), 195–228.
Banerjee, A., Duflo, E., Glennerster, R. and Kinnan, C. (2015). "The Miracle of Microfinance? Evidence from a Randomized Evaluation." American Economic Journal: Applied Economics, 7(1), 22–53.
Roodman, D. (2012). Due Diligence: An Impertinent Inquiry into Microfinance. Center for Global Development.
Yunus, M. (1999). Banker to the Poor. PublicAffairs.
Mills, K. G. and McCarthy, B. (2016). "The State of Small Business Lending: Innovation and Technology and the Implications for Regulation." Harvard Business School Working Paper 17-042.
Mills, K. G. (2024). Fintech, Small Business & the American Dream, 2nd edn. Palgrave Macmillan.
Vasicek, O. (2002). "The Distribution of Loan Portfolio Value." Risk, 15(12), 160–162.
Dunbar, R. I. M. (1992). "Neocortex size as a constraint on group size in primates." Journal of Human Evolution, 22(6), 469–493.
Ostrom, E. (1990). Governing the Commons. Cambridge University Press.
Financial Conduct Authority (2019). PS19/14: Loan-based ('peer-to-peer') and investment-based crowdfunding platforms: Feedback to CP18/20 and final rules.
IFRS Foundation. IFRS 9 Financial Instruments, paragraphs 4.1.1–4.1.4 (the SPPI test).
Lighter Capital. Revenue-Based Financing: terms and repayment caps. lightercapital.com, terms as published.
Kiva. Due diligence and field partner role; published repayment statistics. kiva.org.
Note on figures. Every number above is computed in lib/verify/III_06.py and printed with its inputs, its units and its source by python3 lib/verify.py III.06. Figures marked assumed are assumptions, and each is printed there with its sensitivity. The correlation model follows Vasicek (2002); the normal approximation used for the tails is optimistic, because a real loss distribution is right-skewed.