Do wildfires show up in California's homeless counts?

Eighteen years of federal homeless counts for California's 44 Continuums of Care, compared with state wildfire records, and with estimates from an AI model that never saw the counts.

Daniel Guo · Working draft, updated September 24, 2026 · Code and data on GitHub

Summary

A Continuum of Care (CoC) is the local body that runs HUD's annual Point-in-Time (PIT) count of people experiencing homelessness. This project matches each CoC's yearly count to the wildfires that burned inside its boundary, then asks whether regions with more fire damage count more homeless residents.

Pooled across 756 CoC-years, acres burned and the homeless count are essentially uncorrelated (Pearson r = 0.011). Counts of structures destroyed, available from 2013, give the same result (r = 0.003).
When each region is compared only with its own history, years with larger fire losses have slightly higher counts. None of these estimates reaches statistical significance.
An AI model given only each region's fire history predicted a positive relationship in all 440 of its estimates. On average it overstated the observed correlation by 0.23.
The model's errors were largest when it described itself as highly confident.
Told how the PIT count is collected, the model lowered its estimates but still overstated the correlation. A second model, Gemini, overstated it much less. Neither model ever predicted a negative relationship, and neither could tell which regions had a stronger one.

The PIT count misses many people displaced by fire, such as families staying with relatives or in hotels paid by insurance or FEMA. A weak result here does not show that wildfires have no effect on housing loss. A national study of all 50 states found that homes lost to climate disasters were associated with larger increases in homelessness (Leifheit et al., 2026). The AI's positive direction may be right. This page measures how well it estimates the size of the relationship that appears in HUD's CoC-level counts.

Map

Choose a year and a measure to shade each CoC. The year and measure also apply to the statewide trend and scatter charts below. Click a region, or search for one, to see its history.

Year 2024 Measure Fire perimeters Counts on map

No region selected

Click a region on the map or use the search box.

Source: HUD PIT counts and CoC boundaries; CAL FIRE fire perimeters, fires of 1,000 acres or more. Shading uses a square-root scale so smaller regions remain visible. Boundaries are simplified for display.

Statewide trends

California's total count rose during most of the period. Acres burned by large fires vary much more from year to year, with 2020 standing out.

Figure 1. Homeless count and acres burned in California, 2007–2024

Statewide totals by year. The marked year is the one selected above the map.

Overall homeless

Acres burned in fires of 1,000+ acres

Note: The 2021 drop reflects COVID-19. HUD allowed CoCs to skip the unsheltered count that year, and many California CoCs did. Source: HUD PIT counts; CAL FIRE fire perimeters. Author's calculations.

Fire and homeless counts

Each dot in Figure 2 is one CoC in one year. If fires pushed people into homelessness in the same region, dots further right would tend to sit higher. They do not. Large urban CoCs have the highest counts and rarely burn, while rural CoCs burn the most and have small populations.

Figure 2. Acres burned and homeless count, by CoC and year

Timing Only years with a large fire in the CoC

Note: The line is a least-squares fit. Dots at zero acres are CoC-years with no fire of 1,000 acres or more. Source: HUD PIT counts; CAL FIRE fire perimeters, split between CoCs by area of overlap. Author's calculations.

Pooling all years could hide an effect that appears only in years with severe fires. Figure 3 repeats the comparison one year at a time.

Figure 3. Correlation between acres burned and homeless count, each year

Pearson r across CoCs within a single year, same-year timing.

Note: With 18 tests at p < 0.05, about one false positive is expected by chance. Hover over a bar to see the result with that year's largest-acreage CoC removed. Source: Author's calculations.

Acres burned measures wildland fire, which is only loosely related to lost housing. Table 1 repeats the test with the number of structures CAL FIRE recorded as destroyed. Table 2 uses a stricter design that compares each CoC with its own history and removes statewide changes from year to year.

Table 1. Pooled correlations between fire exposure and the homeless count
Fire measure	Years	Timing	n	Pearson r	Spearman ρ	p

Source: HUD PIT counts; CAL FIRE fire perimeters and Damage Inspection (DINS) records. Author's calculations.

Table 2. Fixed-effects estimates, CoC and year effects removed
Fire measure	Timing	Form	n	Coefficient	Std. error	p

Note: Standard errors are clustered by CoC (44 clusters). In the log form, both variables are log(1 + x), so the coefficient is roughly the percent change in the count for a 1 percent change in fire exposure. In the level form it is additional people counted per acre or per structure. Source: Author's calculations.

AI estimates

An AI model (Claude) received each CoC's name and its year-by-year acres burned, with no homeless counts, and was asked to estimate the correlation between the two. The experiment was run 10 separate times with no memory between runs. Figure 4 compares the average estimate for each CoC with the value computed from the data.

Figure 4. AI-estimated and observed correlation, by CoC

Points on the diagonal would be perfect estimates. Click a point to show that region on the map.

Note: Observed r is the same-year Pearson correlation between acres burned and the overall homeless count within each CoC. Three CoCs with too few fire years are excluded. Source: Author's calculations; estimates from claude-sonnet-5 run through Claude Code.

Figure 5. Average estimation error by the model's stated confidence

Mean absolute difference between each estimate and the observed r, all 10 runs.

Source: Author's calculations.

One possible explanation is that the model leans on well-known recent fires. If so, CoCs whose worst fire was recent should have larger errors. Figure 6 shows no such pattern.

Figure 6. Year of each CoC's largest fire and the size of the AI's error

Source: CAL FIRE fire perimeters. Author's calculations.

Does knowing how the data is collected change the estimates?

Part of the gap could come from the model not knowing how the PIT count works. To test this, the experiment was repeated with one added paragraph describing what the count includes and leaves out. It was also run on a second model, Gemini 3.5 Flash-Lite, with both versions of the prompt. Each of the four combinations had 10 independent runs.

Figure 7. Average AI estimate by model and prompt

Mean estimated correlation across 41 CoCs, compared with the observed average.

Note: The "PIT description" prompt adds a paragraph explaining that the count covers one night in January, includes only people in shelters or unsheltered locations, and skipped unsheltered people in 2021. Gemini 3.5 Flash-Lite is a smaller model than Claude Sonnet; larger Gemini models were not available within free-tier limits. Source: Author's calculations.

Table 4. AI estimates by model and prompt
Model	Prompt	Mean estimate	Overstatement (SE)	Rank agreement ρ	Negative estimates

Note: Overstatement is the mean of (AI estimate minus observed r) across CoCs, with its standard error. Each CoC's observed r rests on about 18 years of data and has a sampling error near ±0.26, so single-CoC comparisons are noisy; the averages are more reliable. Source: Author's calculations.

Show results for every CoC

CoC	Observed r	AI estimate	Error	Confidence

Butte County and the Camp Fire

The 2018 Camp Fire destroyed most of the town of Paradise in Butte County, roughly 14,000 to 19,000 structures. It is the largest loss of housing in this dataset.

Figure 8. Homeless count, Chico, Paradise/Butte County CoC (CA-519)

Overall PIT count by year.

Note: HUD counts only people in shelters or unsheltered on the night of the count. People staying with family, in FEMA housing, or outside the county are not included. The 2021 count excludes unsheltered people because of COVID-19. Source: HUD PIT counts.

Data and methods

Data

Homeless counts: HUD Point-in-Time counts by CoC, 2007–2024, from HUD User.
CoC boundaries: HUD eGIS Open Data.
Fire perimeters: CAL FIRE historical fire perimeters, limited to fires of at least 1,000 acres from 2000 on.
Structures destroyed: CAL FIRE Damage Inspection (DINS), 2013 on.
State outline, used to clip fires and boundaries to California: Census TIGERweb.

Matching fires to regions

Each fire perimeter is intersected with every CoC boundary. When a fire crosses a boundary, its acres are divided in proportion to the area on each side. Each destroyed structure is assigned to the CoC that contains it.

Statistics

Pooled results use Pearson and Spearman correlations across all CoC-years. The yearly results use a Bonferroni correction for 18 tests and a leave-one-out check that drops the CoC with the most acres burned that year. The fixed-effects models include a separate intercept for every CoC and every year, with standard errors clustered by CoC.

AI experiment

The model saw each CoC's name, number, and acres burned by year. It never saw homeless counts or any of the results on this page. It returned an estimated correlation, a direction, a confidence level, and a short explanation for each CoC. All 44 CoCs were sent together in each of 10 independent runs. The estimate shown for a CoC is the mean across runs. The follow-up runs used the same prompt with one added paragraph describing the PIT count, and were repeated on Gemini 3.5 Flash-Lite through the Gemini API. The exact prompts and every raw answer are in the data files below.

Limitations

The analysis is correlational and cannot establish cause in either direction.
PIT counts are a single-night snapshot, and counting methods differ across CoCs and years.
Fires under 1,000 acres are excluded, and CoC boundaries are treated as fixed across all years.
Two AI models were tested, and all regions were estimated in the same request for each run.
The original prompt lists ways fires can cause housing loss and asks for an estimate that is "not a hedge toward zero." This wording may push every estimate upward. A neutral prompt has not been tested yet.

Claude Code was used to help write the data pipeline and this page.

Downloads and citation

All data files used on this page:

Suggested citation

Guo, D. (2026). Do wildfires show up in California's homeless counts? [Working draft and data]. https://research.danielguo.xyz/