Executive summary
Research question: If government funding falls, how many simulated charities move into an annual deficit, how large are the financial gaps, and how would different distributions of extra support grants change those outcomes?
We used seven years of financial information that charities submitted to the Australian Charities and Not-for-profits Commission (ACNC), covering 2018–2024. After screening the records, 41,077 charity profiles were eligible to provide starting financial positions. We selected 10,000 distinct charities from those profiles, created one financial agent for each, and followed their possible finances for three projected years. We repeated the financial simulation five times using the same 10,000 charities. No charity appeared twice within a run.
An annual deficit means that recorded expenses exceed the income recorded for that year. It is not the same as running out of cash or closing. This study did not calculate whether usable reserves could cover a gap, when bills would become due, how leaders would respond, or how spending changes would affect services. Annual public accounts alone cannot answer those questions.
The no-cut comparison means government funding was never reduced by the experimental cut. It does not mean funding was cut and then restored. Even without that cut, simulated income did not cover expenses for approximately 36–37% of charities in year three. Maintaining a 15% cut in each projected year raised that share to approximately 46%. These are annual results in year three, not a count of charities in deficit for three consecutive years.
The main findings were:
- Under a 15% government-revenue cut, roughly 9–10 additional charities in every 100 moved into annual deficit by projected year three, compared with the same financial paths without a cut. This is a simulated accounting result, not a prediction that one in ten real charities will close.
- Starting dependence on government funding mattered. A cut to one revenue stream directly affects organisations that depend more heavily on it.
- Extra grants of equal dollar amounts helped the most agents move from deficit to break-even or surplus. These grants were added after the cut; they were not the continuation of original grants.
- Extra grants based on starting government funding generally reduced the largest total dollar gap at the middle and larger support budgets. They concentrated more money in larger recipients without necessarily eliminating each recipient's deficit.
- More support improved accounting results, but the extra benefit depended on what was measured. Avoiding many small deficits and reducing a few large financial gaps are different objectives.
The experiment is useful for comparing those trade-offs under stated assumptions. It has not established which funding policy preserves the most services, jobs or organisations.
Contents
1 What information was available
The foundation was financial information that charities reported to ACNC for 2018–2024. The records included annual revenue, government revenue, other income and expenses. Source files also contained additional information, including some balance-sheet figures. Those figures were not turned into a validated forecast of available cash or ability to pay bills.
The records were screened for reporting periods, financial consistency, duplicate identities and group reporting. A synthetic demonstration record was excluded. Some unknown reporting information remained explicitly unknown. The 41,077 eligible 2024 profiles were the research population, not all charities and not the whole Australian nonprofit sector.
The study could use these data to ask whether projected annual income covered projected annual expenses. It did not have the linked information and tested rules needed to determine:
- how much cash was available for unrestricted use in each month;
- which reserves or grants were restricted to particular purposes;
- when wages, rent and other bills would fall due;
- what a charity's leaders would choose to change after losing funding;
- how those choices would affect staffing, service volumes, beneficiaries or quality.
For that reason, a deficit in this report cannot be translated directly into closure, job losses or a percentage reduction in service footprint.
2 What the 10,000 agents were
The program did create 10,000 agents. Each was a separate computer record representing a nonprofit financial position. Every record had its own revenue and expense amounts and its own subsequent financial path. This is why it is reasonable to describe them as agents at the accounting level.
Those agents were created within one Python program running on the DGX Spark. They were not chatbot accounts, and did not require separate computers. Representing an organisation as a record is a computational simplification; it is not a compression of a real charity into a fully realistic digital replica.
The agents did not communicate, compete for grants or make leadership decisions. Each was subjected to the same type of funding stress, with its own financial position determining the accounting result. The more precise description is an organisation-level financial simulation. It is not yet a simulation of interacting charity leaders or of the whole nonprofit ecosystem.
One agent for each of 10,000 distinct charities
This version selected 10,000 different source charities without replacement: once selected, a charity could not be picked again. The same starting cohort was used in all five repetitions. Each repetition generated another possible financial path for those charities; it did not add another 10,000 different charities.
The earlier version allowed repeated profiles and varied the starting population between runs. This report uses the completed unique-charity experiment instead. Its results describe one fixed selection from the eligible population. Variation across its five runs reflects simulated financial changes, not uncertainty about which charities were selected. A different selection of 10,000 charities could produce different results.
Why use 10,000 records
A large set of records can retain differences between small and large budgets, government dependence and starting surpluses. It lets the model show a distribution of outcomes rather than treating every charity as an average organisation.
However, 10,000 is a study size, not a threshold at which a model becomes realistic. More agents reduce some sampling noise, but do not repair incorrect assumptions, missing behaviour or incomplete data.
3 How a funding cut was applied
The percentage cut affected government revenue, not total income. Expenses and other financial components followed the same path they would have followed without the cut. The charity was not assumed to replace the lost revenue or reduce its spending in response.
For example, a charity with $1 million in revenue, of which $600,000 came from government, would lose $90,000 under a 15% government-revenue cut. Its total revenue would fall by 9%, before any other financial change.
For each projected year, the model first calculated a possible financial position without the cut. It then reduced that year's government revenue by the chosen percentage. These paired versions made it possible to isolate the accounting effect of the cut.
Two parallel scenarios, with no restoration of cut funding
For every charity, the program compared two versions of the same possible financial future:
- No-cut scenario: the experimental reduction is never imposed. Government revenue may still change according to the historical financial patterns, but no additional 15% cut is applied.
- Continuing-cut scenario: government revenue is 15% lower than the corresponding no-cut amount in each of the three projected years. The reduction remains in place; it is not restored in years two or three.
What a continuing 15% cut means
Suppose government revenue without a cut would be $100,000, $110,000 and $120,000 over three years. With a continuing 15% cut, the corresponding amounts are $85,000, $93,500 and $102,000.
The percentage is applied to each year's otherwise simulated government revenue. It is not a further 15% reduction on the already-cut amount each year. Nor is it a fixed annual dollar reduction calculated only from the starting allocation. Those would be different experiments.
4 How historical patterns were used and checked
Two kinds of historical pattern were needed. One described changes in the amounts of revenue, income and expenses. The other described movements between reporting no government revenue and reporting some government revenue.
The financial-change examples came from adjacent annual records with later years of 2019–2021. Both versions of the forward experiment used that same older financial-change bank.
We then tested two alternatives for the government-funding movements:
| Plain English label used below | What it means | Earlier technical label |
|---|---|---|
| Earlier funding pattern | Uses movements between zero and positive government revenue found among compatible examples ending in 2019–2021 | Pooled |
| 2024 funding pattern | Uses the frequency of those movements in 2024 records, adjusted for the charity's starting characteristics | Recent |
Each pattern was held fixed for the three projected years. The 2024 version did not replace all older financial examples with 2024 examples. It changed the rule for movement into or out of reported government revenue. Neither version proves that future funding will behave that way.
What the historical backtests were
A backtest asks whether the method can reproduce a period for which we already have reported outcomes. One test started from the eligible charities' 2021 accounts, generated their possible finances for 2022–2024, and compared them with their actual reported figures. It used 31,139 charities with complete adjacent reports across those years.
For the version using the most recent funding pattern allowed at that historical starting point, the results were:
| Year being checked | Simulated annual deficit rate | Reported annual deficit rate |
|---|---|---|
| 2022 | 34.18% | 36.72% |
| 2023 | 35.42% | 36.46% |
| 2024 | 36.37% | 36.45% |
The model came close to the final overall deficit rate, but missed important changes underneath it. In 2022, for example, it simulated 13.62% of agents starting with government revenue moving to zero government revenue. In the reported records, the comparable figure was 23.22%—a miss of 9.60 percentage points. This was a change in reported funding, not proof that those charities closed.
Another test restarted from actual prior-year records every year. That can improve a one-year comparison, but does not demonstrate an accurate three-year forecast made at the beginning.
These tests reveal why the results are conditional stress scenarios. Historical years were examined repeatedly during development, the complete-report sample excludes missing reports, and reporting-year labels do not reconstruct exactly when information became available. Agreement with some past totals is not proof of accurate future behaviour.
5 How many more agents entered deficit
The model can produce deficits even when the experimental government cut is never imposed. Some charities already started in deficit; others moved into deficit because their simulated income and expenses changed. The no-cut scenario is therefore a comparison without the extra shock, not a promise of balanced accounts or unchanged revenue.
In year three, approximately 3,590–3,727 of the 10,000 charities had an annual deficit without the experimental cut. With the continuing 15% cut, approximately 4,583–4,590 did. The difference is approximately 856–1,000 additional annual deficits, depending on the funding-pattern assumption.
This raises the deficit rate from approximately 36–37% to 46%, an increase of 8.56–10.00 percentage points. Relative to the original no-cut deficit count, that is about 23–28% more. It does not mean all those charities had deficits in every year, and it does not mean their funding was restored.
| Year-three comparison, five-run average | Earlier funding pattern | 2024 funding pattern |
|---|---|---|
| Government cut never imposed: charities in deficit | 3,590 (35.90%) | 3,727 (37.27%) |
| 15% cut maintained each year: charities in deficit | 4,590 (45.90%) | 4,583 (45.83%) |
| Additional deficits attributable to the cut within the model | 1,000 | 856 |
For example, a charity with $100,000 of annual income and $105,000 of expenses has a $5,000 deficit before any experimental cut. If $40,000 of its income comes from government, a 15% cut removes another $6,000 and deepens its deficit to $11,000. It remains one deficit charity in both counts. A different charity with a $4,000 surplus would cross into deficit after the same $6,000 loss and would count as one additional deficit. These are illustrative examples, not study averages.
Comparing larger funding cuts
The completed unique-charity run included 0%, 10%, 15% and 25% government-revenue cuts. In response to the question about a 50% cut, an additional calculation applied that cut to the same saved no-cut financial paths. It introduces no new management response or financial-path draws.
The 50% row is a new accounting sensitivity, not part of the completed Spark run. The same calculation reproduced the unique-run 15% and 25% counts before adding it.
| Cut to each projected year's government revenue | Earlier funding pattern: year-three deficit rate | Additional deficit agents versus no cut | 2024 funding pattern: deficit rate | Additional deficit agents versus no cut |
|---|---|---|---|---|
| No cut | 35.90% | 0 | 37.27% | 0 |
| 15% | 45.90% | 1,000 | 45.83% | 856 |
| 25% | 50.07% | 1,418 | 49.21% | 1,194 |
| 50% — new sensitivity | 55.50% | 1,960 | 53.34% | 1,607 |
All figures are five-run means; agent counts are rounded. These are percentages of simulated agents, not estimates of Australian charity closure rates.
The count does not double when the cut doubles. Some agents are already in deficit, some have no government funding, and some have enough surplus to remain above zero. A larger cut can also deepen an existing deficit without changing the count at all. This is why the dollar gaps also matter.
6 How large the new deficits were
We separately examined agents that were at break-even or in surplus on the no-cut year-three path, but moved into deficit with the 15% cut.
| Annual deficit among these newly affected agents | Earlier funding pattern | 2024 funding pattern |
|---|---|---|
| Median gap | $89,694 | $114,787 |
| Mean gap | $756,246 | $866,081 |
The median is the middle result: approximately half the observed agent occurrences had a smaller gap and half a larger gap. The much larger mean reflects a small number of very large gaps. These amounts pool simulated outcomes across five repetitions of the same unique cohort. A charity can contribute an outcome in more than one repetition; these are not 50,000 different charities or a national census.
A gap of $90,000 could be manageable for one organisation and critical for another. To establish that difference, a cash forecast would need to track usable money, payment timing and restrictions. A service model would need to connect staffing, fixed costs, contracts and spending to service delivery. Neither was built in this accounting experiment.
7 Funding dependence and diversification
The model distinguished government revenue from combined nongovernment revenue. It did not measure whether nongovernment income came from many independent sources or one large funder. Therefore, the comparison below measures dependence on government, not complete funding diversification.
For the 15% cut, the increase in year-three deficit rates was:
| Government share of revenue at the start | Earlier funding pattern | 2024 funding pattern |
|---|---|---|
| More than zero but below 25% | +8.42 percentage points | +7.14 points |
| 25% to below 75% | +23.28 points | +20.93 points |
| At least 75% | +37.61 points | +36.26 points |
Some agents starting with no government revenue acquired it along their simulated path and were therefore affected later. Their incremental rates were 2.38 and 1.07 points.
Part of the relationship is direct arithmetic. A charity obtaining 20% of revenue from government loses 3% of total revenue under a 15% government cut; one obtaining 80% loses 12%, before any response. Size, starting margins and funding dependence also overlap. The experiment does not establish an independent causal effect of diversification or show that leaders successfully found replacement funding.
8 What extra support grants meant
The support experiment asked a second question: After the 15% cut, how would a separate budget of additional annual grants change the accounting results?
The original government funding was still cut. Support was then added as unrestricted income. It was not continuation of the old grant, and was not automatically matched to each agent's lost funding or future deficit.
Each allocation rule received the same total budget. Grants were calculated from starting information and held fixed for the three projected years. Expenses and financial-path decisions were unchanged.
| Allocation rule | Plain English explanation | Example at the middle support budget |
|---|---|---|
| Equal grants | Every agent receives the same additional dollar amount | Approximately $49,145 per agent annually, whether large or small |
| Government-proportional grants | Each agent's share of support follows its share of starting government revenue | An agent with $1m starting government revenue receives $37,500 annually; one with $10m receives $375,000 before caps |
| Thin-margin targeted grants | Proportional grants, but only for agents starting with government share of at least 25% and a surplus margin from zero to below 10% | The eligible group shares the entire support budget |
| Grant caps | Set a maximum individual grant and redistribute the excess within the same eligible group | Less concentration in the largest recipients |
For the proportional rule at the middle budget, each uncapped grant equals 3.75% of starting government revenue: 25% of a 15% starting-year loss. That is not necessarily 25% of a later projected year's loss, because its government revenue may have changed.
Why different definitions of success favour different choices
| If success means | Measure needed | What this experiment can say |
|---|---|---|
| Help the largest number reach break-even | Number of annual deficits avoided | Equal grants perform best among tested rules |
| Reduce the largest combined financial gap | Total deficit dollars reduced | Government-proportional grants generally lead at the middle and larger budgets |
| Focus on damage specifically caused by the cut | Deficits avoided that did not exist without the cut | Reported separately; do not count all deficits avoided as shock relief |
| Reduce repeated financial stress | Agents avoiding consecutive projected deficit years | Measured, but not proof of survival |
| Limit funding concentration | Share of support going to the largest recipients | Caps spread allocations more broadly |
| Preserve jobs, essential services, beneficiaries or organisational survival | Validated staffing, service, cash and response measures | Not answered by the current study |
At the middle budget, year-three results were:
| Rule | Annual deficits avoided earlier / 2024 funding pattern | Deficit dollars reduced earlier / 2024 funding pattern |
|---|---|---|
| Equal grants | 2,580 / 2,542 | $132.3m / $132.8m |
| Government-proportional grants | 158 / 158 | $359.0m / $345.1m |
An additional breakdown of the saved records makes the size effect clearer. At the middle budget, approximately 92% of the annual deficits eliminated by equal grants were among agents starting below $500,000 in annual revenue. Approximately 94% of the deficit-dollar relief from proportional grants went to agents starting at $3 million or more.
| Starting annual revenue | Equal grants: annual deficits avoided, earlier pattern | Proportional grants: deficit dollars reduced, earlier pattern |
|---|---|---|
| Below $500,000 | 2,372 | $2.3m |
| $500,000 to below $3m | 185 | $18.5m |
| $3m or more | 23 | $338.2m |
These are five-run means for year three; the 2024 funding pattern gives similar proportions. The figures group agents by observed starting revenue, not by their later projected size. This is an additional breakdown of the verified unique-charity trajectories, retained separately from the completed Spark outputs.
Proportional grants therefore concentrated financial relief in larger starting organisations. That does not establish that their services benefited more, or that each large organisation benefited more than each small one. The largest starting 1% received 45.35% of proportional support, compared with 1% of equal support.
A large grant can reduce a large deficit without eliminating it. A smaller equal grant can eliminate a small deficit completely. Counting only organisations crossing zero would overlook the first benefit; counting only dollars would overlook the breadth of the second.
9 How support budgets and grant caps changed the result
There were two separate percentages in these experiments:
- Funding-cut percentage: how much government revenue was removed. Section 5 compares 15%, 25% and the new 50% shock.
- Support-budget percentage: how much additional support was made available after the 15% shock. The support grid was 10%, 25% and 50% of the starting-year cut amount.
They should not be read as the same scenario. In the support grid, the underlying government-revenue cut remained 15%.
The support budget was calculated from observed starting government revenue, then fixed across projected years. Mean budgets and equal-grant results were:
| Support budget as share of starting-year cut | Mean annual support | Equal annual grant per agent | Year-three annual deficits avoided earlier / 2024 pattern |
|---|---|---|---|
| 10% | $196.6m | $19,658 | 1,914 / 1,908 |
| 25% | $491.5m | $49,145 | 2,580 / 2,542 |
| 50% | $982.9m | $98,291 | 3,084 / 3,029 |
Moving from 10% to 25% made the support budget 2.5 times larger. Moving from 25% to 50% doubled it. That doubling avoided another 505 or 487 annual deficits under equal grants: approximately 20% or 19% more, rather than twice as many. Those counts include some deficits that would have existed even without the cut.
At the largest budget, uncapped proportional grants reduced about $657–685m of deficit dollars. A fixed dollar cap spread support more broadly and avoided more individual deficits, but reduced less of the aggregate gap. No allocation rule became a universal winner.
Money above a recipient's same-year accounting deficit was reported separately. It could have other value—such as building a buffer—but that value was not modelled. It should not automatically be labelled waste.
10 What a future simulation of leaders' decisions would require
Real charity leaders do not passively accept unchanged spending after funding falls. They may use available reserves, seek donations, apply for grants, raise fees, reduce hours or staffing, change services, collaborate, or withdraw from a programme. The options depend on restrictions, contracts, demand, resources and leadership judgement.
The present model holds those responses outside the calculation so that the immediate accounting exposure is visible. It shows the consequence before an explicit response model, not the inevitable real-world outcome.
A future behavioural simulation could add evidence-based descriptions of organisations and decision-makers. A useful persona would describe things that affect feasible choices: reliance on particular contracts, usable reserves, fundraising staff, access to alternative funders, service obligations, implementation delays and appetite for financial risk. It should not simply assign a personality label and assume that label predicts success.
The development path would be:
- Collect examples of what comparable charities actually did after funding changes, using linked records, interviews, surveys and operational information.
- Define possible actions, eligibility, timing and costs. For example, fundraising can take months, incur costs and fail; staffing reductions can face contractual constraints.
- Estimate and test response probabilities for different circumstances. Risk appetite should interact with resources and obligations, rather than overriding them.
- Connect actions to cash, expenses and service measures, including the possibility that a financially helpful response harms service access.
- Test the model against new cases not used to design it, and show sensitivity when behavioural evidence is weak.
A language model could help express or explore decisions, but plausible explanations are not evidence that those decisions occur at the assumed frequency. Behavioural rules still need observation and validation. Interactions between charities, funders and beneficiaries would require further work and could change aggregate outcomes. For example, if many charities seek replacement donations at the same time, they may compete for the same limited funding; a realistic response model cannot assume that every fundraising attempt succeeds independently.
11 Conclusion
The study created one financial agent for each of 10,000 distinct charities to preserve differences between organisations and compare their accounting exposure under common stresses. This is a useful representation of part of the real world: diverse organisations facing different financial gaps. It does not recreate the full environment in which leaders make decisions.
The evidence supports three conclusions within the tested assumptions: larger government-revenue cuts increase annual deficits; dependence on that revenue stream increases exposure; and support allocation involves a trade-off between the number of deficits eliminated and total deficit dollars reduced.
It does not support conclusions about how many charities would close or halve their services. Adding observed behaviour, timing, cash restrictions and service relationships could make the simulation answer those broader questions. Until then, it is a transparent accounting stress test and a basis for further research.
Notes on repetition, evidence and the new sensitivity
The starting cohort was selected once using seed 42. Financial replications were labelled 42–46, with transition generators using those numbers plus 1,000,000. Seeds control repeatable randomness; they are not economic assumptions. Within each funding pattern, cut and support comparisons shared the same underlying financial paths.
For the 15% cut in year three, additional deficits ranged from 980–1,022 across the five earlier-pattern runs and 836–871 across the 2024-pattern runs. These ranges measure financial simulation variability conditional on this fixed cohort and model. They do not measure variation across alternative charity selections or provide confidence bounds for the real world.
This report uses the completed unique-charity run, verified against inventory SHA-256 4186faf172f6f5760197dcd0919620a3a650034ed66a5a785b3fad312e7f237f. All 176 inventoried files matched their hashes. The sampled identifiers confirm 10,000 distinct charities in each repetition and the same starting cohort across seeds. The historical backtest results are unchanged from the earlier study because they use the same historical evaluation cohorts and methods.
Additional local analyses for this report are retained in shock-extension/, deficit-gap-analysis.json and support-size-analysis.json. The 50% cut is an additional accounting overlay on the saved unique-cohort paths, not a new Spark simulation. The original completed outputs and previous reports remain unchanged. All figures in this report refer to the unique-charity experiment or explicitly identified historical checks and illustrative examples.
Data context: ACNC, 2024 Annual Information Statement Guide, https://www.acnc.gov.au/tools/guides/2024-annual-information-statement-guide . This report explains the operator's verified outputs and additional accounting analyses; it is not independent validation of forecast or causal accuracy.