From Economic Indicators to a Practical Predictive Model
September 2026
Executive Summary
Can Canadian consumer insolvencies be forecast with reasonable accuracy?
That question was the starting point for an extensive analysis of Canadian consumer insolvency statistics and economic data. Historical insolvency filings published by the Office of the Superintendent of Bankruptcy Canada (OSB) were combined with population and economic data from Statistics Canada. More than 40 regressions were examined, including different economic variables and time lags.
The initial hypothesis was straightforward. Consumer insolvencies should be influenced by economic conditions such as household debt, debt-service costs, unemployment, inflation, household income, wages and interest rates. It therefore seemed reasonable that some combination of these variables might provide a useful method of forecasting future consumer bankruptcies and proposals.
The research produced a somewhat different result.
The single most useful predictor of Canada’s consumer insolvency rate in the next quarter was not an external economic variable. It was Canada’s insolvency rate in the previous quarter.
Canadian consumer insolvency rates exhibit substantial persistence. A relatively high rate in one quarter tends to be followed by a relatively high rate in the next. A relatively low rate tends to be followed by another relatively low rate.
A second factor materially improves the forecast: quarter-of-year seasonality.
Forecast insolvency rate = previous quarter’s insolvency rate + quarter-of-year seasonality
The forecasting process operates in three stages: the regression predicts the quarterly consumer insolvency rate per 1,000 Canadians aged 15 and over. Then the predicted rate is applied to the population aged 15 and over to obtain a forecast of the actual number of consumer insolvency filings. Finally, the quarterly forecast is allocated among the three months using the average monthly filing distribution observed over the most recent ten years.
Current Forecast: 2026 Q3
| Forecast quarter | 2026 Q3 |
| Forecast filings | 35,604 |
| Forecast rate | 1.0221 per 1,000 |
| Previous quarter actual | 37,519 |
| Forecast change from 2026 Q2 | −5.1% |
| Month | Forecast |
| July | 11,408 |
| August | 11,825 |
| September | 12,371 |
| Total | 35,604 |
Why Forecast Consumer Insolvencies?
Licensed Insolvency Trustees encounter financial distress at the individual level.
A debtor may lose employment, experience reduced income, accumulate excessive credit-card or other unsecured debt, face higher mortgage or loan payments, experience marital breakdown or simply reach the point where existing financial obligations can no longer be serviced.
At the national level, thousands of these individual circumstances eventually appear in the administrative statistics collected by the Office of the Superintendent of Bankruptcy. Those statistics create a long historical record of Canadian consumer bankruptcies and consumer proposals.
That historical record raises an interesting question: how much information about future consumer insolvencies is contained in the data already available today?
Forecasting has several practical uses.
It provides context for new OSB statistics. Rather than simply observing that insolvencies rose or fell during a quarter, an actual result can be compared with what historical behaviour suggested was likely.
Forecasting can also help identify unusual periods. If actual filings consistently depart from the model’s expected range, that may indicate that something important has changed in the economy, consumer behaviour, government policy, creditor practices or the insolvency system.
Most importantly, forecasting imposes statistical discipline on economic explanations.
It is easy to state that high interest rates, unemployment or household debt should increase consumer insolvencies. Those propositions may be economically sensible.
A different and more demanding question is: does including that information actually make forecasts more accurate?
That distinction became one of the central themes of this research.
The Data
The dependent-variable data are Canadian consumer insolvency filings reported by the Office of the Superintendent of Bankruptcy Canada. The historical database used here extends back to the late 1980s and includes consumer bankruptcies and consumer proposals.
Readers can explore Canadian consumer insolvency trends by province, city, CMA and forward sortation area using the Interactive Consumer Insolvency Data Explorer.
Economic and population variables examined during model development were obtained primarily from Statistics Canada.
The forecasting program also performs a series of validation checks before estimating the model. Among other things, it checks that quarterly filing, population and monthly filing data are present; that quarterly and monthly dates are not duplicated; that quarters are not missing; and that the population series covers the latest actual insolvency quarter.
These controls are important because forecasting errors can arise not only from a poor statistical model but also from incomplete, duplicated or incorrectly aligned data.
What Should Be Forecast?
The earliest regressions run used the actual number of consumer insolvency filings.
Figure 1. Quarterly Canadian consumer insolvency filings and population aged 15+, 1987 to 2026 Q2.
The long-term growth in Canada’s population illustrates why raw filing counts can be misleading when comparing insolvency activity across several decades.
The number of insolvency filings seems like the obvious dependent variable. If the objective is to forecast how many bankruptcies and proposals will be filed, why not model the number directly? The problem is population growth.
Canada has substantially more people today than it did in the late 1980s or 1990s. Even if the individual probability of insolvency remained unchanged, a larger population could produce a larger number of insolvency filings. Raw filing counts therefore contain at least two components: changes in the underlying rate of consumer insolvency and changes in the size of the Canadian population. This can also distort regressions involving economic variables that trend over time.
The solution was to model a population-adjusted insolvency rate.
Quarterly consumer insolvency rate = quarterly filings ÷ population aged 15+ × 1,000
Why Population Aged 15 and Over?
The population aged 15 and over is used as a consistent population denominator and is available from Statistics Canada over the required historical period. It is not intended to imply that every person aged 15 or over is equally likely or legally positioned to make an insolvency filing.
Its role is statistical: it provides a consistently measured adult-oriented population series with which to adjust filing counts over time. The important consideration is consistency. Once the dependent variable is expressed as an insolvency rate, changes in Canada’s population are substantially separated from changes in the propensity for consumer insolvencies to occur.
Population Is Not an Independent Variable in the Regression
Population remains important, but its role should not be confused with the variables in the forecasting equation. The regression forecasts the insolvency rate. Population is then used to convert that predicted rate into a forecast number of filings.
Forecast filings = forecast insolvency rate × population aged 15+ in thousands
This creates a useful separation. The regression estimates how prevalent insolvency filings are expected to be. Population determines how many people that rate applies to.
This also avoids asking the regression to distinguish population growth from other long-term trends.
The Economic Variables Investigated
The research began with a much broader model-selection exercise than the final equation might suggest.
Economic variables examined included:
- household debt-service ratio;
- unemployment;
- Consumer Price Index and inflation;
- household disposable income;
- household credit-market debt;
- non-mortgage debt;
- wages;
- household debt relative to disposable income; and
- interest rates.
These economic relationships are examined in more detail in What Drives Consumer Insolvencies in Canada?
Each has a plausible relationship with household financial distress. The debt-service ratio measures the portion of household disposable income required to service debt. As that burden rises, households have less financial flexibility. Unemployment can cause an immediate loss of income and may eventually make existing debts unsustainable. Inflation can reduce the amount of income available after necessities are purchased. Interest rates affect mortgage, line-of-credit and other borrowing costs. Disposable income and wages affect households’ ability to meet existing obligations. Household debt measures the amount of indebtedness that ultimately has to be serviced. Non-mortgage debt is of particular interest in consumer insolvency because unsecured debt frequently represents a substantial portion of the claims dealt with through bankruptcies and consumer proposals.
There were therefore sound reasons to investigate all of these variables.
But economic plausibility and forecasting usefulness are not the same thing.
Simple Regression as the Starting Point
The initial analysis relied heavily on simple regressions. A simple regression examines the statistical relationship between one explanatory variable and the dependent variable.
For example: is a higher household debt-service ratio associated with a higher consumer insolvency rate?
The regression produces a coefficient estimating the direction and size of the relationship and statistics describing how well the model fits the historical observations.
One commonly reported measure is R-squared. R-squared measures the proportion of historical variation in the dependent variable that is accounted for by the regression. An R-squared of 0.50 means that approximately half of the historical variation is accounted for by the fitted relationship.
But R-squared has important limitations. It does not establish causation. It does not necessarily indicate that the relationship will remain stable. And, crucially for this project, it does not demonstrate that the model can forecast observations it has never seen.
The Debt-Service Ratio: A Useful Warning
The household debt-service ratio illustrates why caution is necessary.
When raw quarterly insolvency filing counts were initially used as the dependent variable, the debt-service ratio appeared to have a relatively strong relationship with insolvencies. The R-squared was roughly in the 0.66 to 0.70 range.
That initially appeared promising.
After the dependent variable was changed to the population-adjusted insolvency rate, however, the relationship became considerably weaker. The same-quarter debt-service-ratio regression produced an R-squared of approximately 0.325. The results at lags of one through four quarters were also around 0.30 to 0.32.
This does not mean that debt-service burdens are irrelevant to household insolvency. Rather, it demonstrates that part of the impressive-looking relationship with raw filing counts was associated with broader historical trends. Population adjustment helped reveal that.
This was an important methodological step in the project.
Why Economic Variables Were Lagged
Financial distress often develops gradually.
Suppose interest rates rise. A household may initially reduce discretionary spending. It may then use savings or additional credit. Later it may miss payments. Collection activity may begin. Only after those stages might the debtor seek professional advice and file a bankruptcy or consumer proposal.
The same reasoning applies to unemployment, reductions in income and other economic shocks.
Consequently, an economic variable may have little relationship with insolvencies in the same quarter but a stronger relationship several quarters later.
The analysis therefore tested different lags, generally from zero to four quarters. This substantially increased the number of regressions examined. More than 40 specifications were considered during the exploratory stage.
That exercise was useful for understanding the data, but it also increased the danger of selecting a model simply because one specification happened to fit the historical sample particularly well. This reinforced the need for genuine out-of-sample testing.
The Strongest Predictor Was Already in the Insolvency Data
The major change in direction occurred when the previous quarter’s insolvency rate was introduced as an explanatory variable. It dominated the analysis.
In the current fitted model, the coefficient on the previous quarter’s insolvency rate is 0.948822, with a t-statistic of approximately 46.20. The relationship is extremely strong statistically.
In practical terms, if the previous quarter’s insolvency rate is 0.10 per 1,000 higher, the model predicts that the following quarter’s rate will be approximately 0.095 higher before adjusting for quarter-of-year seasonality.
Canadian consumer insolvency activity therefore exhibits substantial persistence.
Why Persistence Makes Sense
Consumer insolvency conditions do not reset every three months.
The stock of household debt changes gradually. Employment conditions generally evolve over time rather than disappearing at quarter-end. Interest-rate effects can remain in household budgets for extended periods. Financially distressed consumers may take months before deciding to consult a Licensed Insolvency Trustee. Creditor practices and access to credit also change gradually.
The current insolvency rate therefore contains information about many conditions that remain relevant in the next quarter. It effectively summarizes part of the recent economic and financial environment.
This helps explain why the previous quarter’s insolvency rate can outperform individual economic indicators for short-term forecasting.
This Does Not Mean Economic Conditions Do Not Matter
The forecasting result must be interpreted carefully. It would be incorrect to conclude that interest rates, unemployment, debt or income do not affect consumer insolvencies.
The correct conclusion is narrower: once the current insolvency rate is known, the economic variables examined provided relatively limited additional information for predicting the immediately following quarter compared with the predictive information already contained in the insolvency rate itself.
The current rate may already reflect the accumulated consequences of earlier economic developments.
For example, higher interest rates may place households under increasing pressure. Some households subsequently become insolvent. Once that increase begins appearing in OSB filing statistics, part of the interest-rate effect is already embedded in the observed insolvency rate.
This is an important distinction between prediction and causation.
Quarter-of-Year Seasonality
The second component of the final model is quarter-of-year seasonality.
The current regression coefficients are:
| Variable | Estimate | p-value |
| Intercept | 0.071492 | 0.000687 |
| Previous-quarter insolvency rate | 0.948822 | <0.000001 |
| Q2 | 0.000971 | 0.943747 |
| Q3 | -0.073129 | <0.000001 |
| Q4 | -0.012397 | 0.371212 |
Q1 is the reference quarter.
The most substantial seasonal effect occurs in Q3. Holding the previous quarter’s insolvency rate constant, Q3 historically has an insolvency rate approximately 0.073 per 1,000 lower than Q1.
Q2’s estimated difference from Q1 is essentially zero. Q4 is modestly lower, but its individual coefficient is not statistically significant at conventional levels.
Looking only at these p-values could lead to the conclusion that seasonality contributes relatively little. That conclusion would be wrong.
Testing Whether Seasonality Actually Improves Forecasts
Instead of deciding whether to retain seasonality from the individual coefficient p-values, two competing models were tested:
Model A: previous-quarter insolvency rate only. Model B: previous-quarter insolvency rate + quarter seasonality. Both were subjected to the same rolling historical forecasting procedure.
The current results are:
| Measure | Lag 1 only | Lag 1 + Quarter |
| MAE | 0.0527 | 0.0411 |
| RMSE | 0.0759 | 0.0691 |
| Median Absolute Error | 0.0387 | 0.0247 |
| Mean Signed Error | -0.0003 | -0.0018 |
Adding quarter seasonality reduces MAE by 22.06% and RMSE by 8.95%. Median absolute error declines by approximately 36%. That is persuasive forecasting evidence.
Seasonality remains in the model not merely because one quarter has a statistically significant coefficient, but because including the quarter variable produces materially better forecasts of unseen historical observations.
Statistical Significance Is Not Forecasting Performance
This provides one of the clearest lessons from the research. A variable can be statistically significant and still add little practical forecasting value. Conversely, evaluating a group of variables solely from individual p-values can overlook useful predictive information.
For a forecasting model, the most relevant question is not simply whether the coefficient is statistically significant. It is whether the model forecasts better when this information is included.
For quarter seasonality, the answer is clearly yes.
The Production Regression
The current production model is:
Forecast insolvency rate = previous-quarter insolvency rate + quarter-of-year seasonality
Using Q1 as the reference quarter, the fitted equation is:
Forecast Rate(t) = 0.071492 + 0.948822 × Rate(t−1) + 0.000971 × Q2 − 0.073129 × Q3 − 0.012397 × Q4
Only the indicator for the quarter being forecast takes the value 1. The others equal zero.
There is no debt-service ratio in the production equation. There is no unemployment variable. There is no interest-rate variable. There is no CPI variable. Population is not an independent regression variable.
The simplicity of the equation is the result of the model-development process rather than an assumption imposed at the beginning.
Historical Fit of the Current Model
To Q2 2026, the model contains 157 observations.
Its R-squared is 0.9340 and adjusted R-squared is 0.9323.
Approximately 93.4% of the historical variation in Canada’s quarterly consumer insolvency rate is therefore accounted for by the model. The addition of Q2 barely changed the fitted model.
Before Q2 was incorporated, R-squared was 0.9338 and the lagged-rate coefficient was 0.948970. After Q2, R-squared is 0.9340 and the lagged-rate coefficient is 0.948822. The Q3 seasonal coefficient changed from approximately -0.073135 to -0.073129.
This stability is reassuring. A single additional quarter has not materially changed the estimated historical relationships.
Why Historical Fit Is Not Enough
A model can achieve a high R-squared and still forecast poorly. If enough variables are tested, a model can begin fitting historical noise rather than stable relationships. This is generally described as overfitting.
A model designed using the entire dataset also benefits from knowledge of observations that would not have been available had the forecast actually been made at the time.
For these reasons, the production model is evaluated using rolling historical forecasts. This is arguably more important than the model’s 0.934 R-squared.
Rolling Out-of-Sample Validation
The rolling procedure begins with an initial training period of 40 quarters, approximately ten years.
Figure 3. Rolling one-quarter-ahead forecast errors. Positive values mean actual insolvency rates exceeded the forecast; negative values mean the model forecast too high. The zero line represents a perfect forecast.
The model is fitted to those observations. It then forecasts the next quarter. That forecast is recorded. The actual observation is subsequently added to the training data, the model is re-estimated, and the next quarter is forecast.
This process continues through the historical series. With Q2 2026 now incorporated, there are 117 rolling historical forecasts. Each forecast is generated without using the actual value of the quarter being forecast to estimate that particular model.
This is designed to approximate how the forecasting system would have operated in real time.
Mean Absolute Error
The current rolling Mean Absolute Error (MAE) is 0.0411 insolvencies per 1,000 population aged 15+.
MAE measures the average magnitude of forecast errors while ignoring whether each error was positive or negative. It is particularly useful because its units are the same as the insolvency rate being forecast.
A lower MAE indicates better forecasting performance.
Root Mean Squared Error
The current Root Mean Squared Error (RMSE) is 0.0691. RMSE penalizes large forecasting errors more heavily than MAE because errors are squared before averaging.
The fact that RMSE is appreciably higher than MAE indicates that the historical record includes some quarters with unusually large forecasting errors.
That is to be expected in a series extending through major economic and institutional disruptions.
Median Absolute Error
The median absolute error is 0.0247.
Half of the rolling forecasts had an absolute rate error smaller than approximately 0.0247, and half had a larger error.
Because the median is less affected by extreme observations, it provides a useful indication of the error associated with a more typical forecast.
Mean Signed Error
The current mean signed error is -0.0018.
Forecast error is defined as Actual − Forecast. A persistent positive value would indicate that the model generally forecasts too low. A persistent negative value would indicate that it generally forecasts too high.
The result of -0.0018 is very close to zero, indicating little overall historical directional bias.
The First Live Test: Q2 2026
Historical backtesting is valuable, but a genuine live forecast is more convincing.
Before Q2 2026 actual filings were known and incorporated into the production dataset, the model forecast 37,783 filings. The actual quarterly total subsequently incorporated into the production data was 37,519 filings.
With Q2 incorporated into the model, the latest actual information is: latest actual quarter 2026 Q2; actual filings 37,519; actual insolvency rate 1.0789 per 1,000 population aged 15+.
The model therefore overforecast by 264 filings, or approximately 0.70%. This is an encouraging result. The forecast was also comfortably inside both forecast ranges that had been established in advance.
However, one accurate forecast does not validate a forecasting model. A model can produce an excellent forecast by chance. The importance of Q2 is therefore not that it proves the model works. Its importance is that it marks the beginning of a genuine live forecasting record.
The next forecast, and the one after that, can be evaluated in exactly the same manner.
The Current Forecast: Q3 2026
The model now forecasts Q3 2026. The predicted insolvency rate is 1.0221 per 1,000 population aged 15+. The estimated Q3 population aged 15 and over is 34,834,451.
Applying the predicted rate to that population produces a Q3 2026 forecast of 35,604 consumer insolvency filings. This represents a predicted decline of approximately 1,915 filings, or about 5.1%, from Q2.
The decline is largely attributable to the model’s Q3 seasonal effect.
Why the Q3 Forecast Falls
The latest Q2 insolvency rate is approximately 1.0789. The lag coefficient is 0.948822.
Q3 also has a seasonal adjustment of -0.073129.
Consequently, the Q3 equation is approximately: Forecast Q3 rate = 0.071492 + 0.948822 × Q2 rate − 0.073129.
Using the unrounded Q2 rate, the model produces the forecast rate of approximately 1.0221.
The predicted decline therefore does not result from a judgment that the Canadian economy will improve during Q3. It results from the historical persistence of insolvency rates combined with the recurring Q3 seasonal pattern found in the data.
Estimating Q3 Population
When the Q3 forecast was prepared, the Statistics Canada population series available to the model extended only through June 30, 2026. An actual Q3 population observation was unavailable.
The production procedure estimates Q3 population from the average quarterly population growth rate over the most recent four quarterly changes. The resulting average growth assumption is 0.174% per quarter. This produces an estimated Q3 population aged 15 and over of 34,834,451.
Generally, population changes relatively slowly from quarter to quarter, so a modest population-estimation error should have a comparatively small effect on the resulting insolvency filing forecast.
Monthly Q3 Estimates
The quarterly forecast is allocated among July, August and September using recent ten-year historical monthly shares for Q3.
| Month | Historical share | Forecast |
| July | 32.04% | 11,408 |
| August | 33.21% | 11,825 |
| September | 34.75% | 12,371 |
| Q3 Total | 100.00% | 35,604 |
These numbers are not three independent monthly regression forecasts. The regression forecasts the Q3 total. The monthly figures merely allocate that quarterly total according to the average recent historical distribution of filings within Q3.
The quarterly forecast should therefore be regarded as the principal statistical forecast.
Why Use a Ten-Year Monthly History?
Using the complete monthly history back to the late 1980s could give substantial weight to administrative and filing patterns that may no longer be representative.
At the opposite extreme, using only the last year or two could make the allocation excessively sensitive to temporary fluctuations.
A recent ten-year window provides a compromise. It is long enough to average across individual unusual years while remaining more representative of contemporary filing patterns.
The Three-Stage Forecasting System
Stage 1 – Forecast the insolvency rate
Previous-quarter insolvency rate + quarter-of-year seasonality → forecast insolvency rate per 1,000 population aged 15+
Stage 2 – Convert the rate into filings
Forecast insolvency rate × forecast-quarter population aged 15+ → forecast quarterly consumer insolvency filings
Stage 3 – Allocate the quarterly forecast
Quarterly forecast × recent ten-year monthly shares → monthly filing estimates
Each stage performs a distinct function. That makes the model comparatively easy to understand, reproduce and audit.
Why Not Use a Large Multiple Regression?
A model containing unemployment, interest rates, inflation, debt, income and other economic variables might appear more sophisticated. It would not necessarily be better.
Additional variables can increase historical R-squared even when they contribute little useful forecasting information. They can also produce unstable coefficients, multicollinearity and overfitting.
The relevant test is therefore not how impressive the equation looks. It is whether additional complexity reduces forecast errors on observations the model did not see during estimation.
Seasonality passes that test.
The broader economic variables investigated to date have not demonstrated sufficient incremental one-quarter-ahead forecasting value to justify replacing the simpler production specification.
Forecasting and Causation Are Different Questions
The final model is not intended as a causal theory of Canadian consumer insolvency.
The previous quarter’s insolvency rate does not cause the next quarter’s insolvencies. It predicts them.
Similarly, the absence of unemployment or interest rates from the final equation does not mean those variables have no effect on household financial distress.
A causal analysis might ask what is the effect of a one-percentage-point increase in unemployment on future insolvency rates?
This forecasting model asks: given the information available today, what provides the most reliable estimate of next quarter’s insolvency rate?
Those are different research questions. The final equation should be interpreted accordingly.
COVID-19 and Structural Change
No historical forecasting model can be expected to handle unprecedented events perfectly.
The COVID-19 period is an obvious example. Employment patterns changed abruptly. Governments introduced extraordinary income-support programs. Consumer spending changed. Courts, creditors and insolvency practices were disrupted. Household savings and borrowing behaviour changed.
The resulting insolvency experience was difficult to predict from historical relationships.
Periods such as COVID-19 help explain why RMSE is higher than MAE: a relatively small number of large errors can materially affect the squared-error measure. They also illustrate why the model must continue to be monitored.
A relationship that worked well historically may weaken if the institutional or economic environment changes.
Avoiding Look-Ahead Bias
A subtle risk in forecasting research is accidentally using information that would not have been available when a historical forecast was supposedly made.
The rolling validation procedure is designed to reduce this problem. At each stage, the model is estimated using only earlier quarterly observations. The next quarter is then forecast. Only after that forecast is recorded is the actual observation effectively added to the expanding training sample.
The same principle is applied to the monthly allocation data: monthly history is restricted to completed information available through the latest actual quarterly period.
This makes the backtesting exercise much closer to genuine prospective forecasting.
Data Quality Is Part of the Model
Forecasting is not simply a matter of writing a regression formula.
The production program performs extensive validation before forecasting. It checks for missing data, duplicate dates, missing quarters and population coverage. It also distinguishes between actual and estimated forecast-quarter population.
If critical information is unavailable or inconsistent with the required model structure, the program is designed to stop rather than silently continue. This is important because a technically correct regression operating on incorrect data can produce a highly precise but completely wrong result.
The data-processing system is therefore part of the forecasting methodology, not merely a preliminary administrative task.
Building a Live Forecast Record
The Q2 2026 result should be treated as Forecast No. 1 in a permanent live forecasting record.
A simple table can eventually contain:
| Forecast quarter | Forecast | Actual | Difference | % error |
| 2026 Q2 | 37,783 | 37,519 | +264 | +0.70% |
| 2026 Q3 | 35,604 | Pending | Pending | Pending |
Over time, this table will become increasingly important.
Historical rolling validation provides 117 simulated out-of-sample forecasts.
The live record provides something different: forecasts genuinely made and recorded before the results were known. After several years, the live record may become the strongest evidence for—or against—the continued usefulness of the model.
The Model Should Not Be Changed Every Quarter
An unexpectedly high or low result should not automatically trigger a model change.
Every forecasting model makes errors.
Changing the specification after every miss would risk fitting random fluctuations rather than improving genuine predictive ability. Model changes should instead require evidence.
For example, a new economic variable might be added if it consistently and materially reduces rolling out-of-sample errors. A model might also require reconsideration if mean signed errors begin showing persistent bias or if MAE and RMSE deteriorate substantially over time.
The existing model now provides a benchmark against which future alternatives can be tested.
Future Research
There are several useful directions for additional work.
Economic variables that add little to a one-quarter-ahead level forecast may prove useful in forecasting turning points. They may also become more important at longer horizons, where the previous quarter’s insolvency rate provides less information.
Provincial models could reveal different persistence or seasonal characteristics.
Non-mortgage debt deserves continued investigation because of its close conceptual relationship with consumer insolvency.
Alternative time-series models could also be tested against the existing production model.
But any alternative should face the same standard: does it materially improve out-of-sample forecasting performance? Complexity alone is not an improvement.
What This Means for Licensed Insolvency Trustees
The most interesting conclusion for insolvency professionals may be the degree of momentum in national consumer insolvency activity. Current filing activity contains a remarkable amount of information about the immediate future.
That does not make economic conditions unimportant.
Rather, the current insolvency rate appears to summarize much of the recent economic and financial pressure that has already progressed far enough to affect actual filings. For a one-quarter-ahead national forecast, that information is extraordinarily valuable.
Quarter-of-year seasonality then provides an additional adjustment that has demonstrated substantial improvement in historical forecasting accuracy.
The result is a model considerably simpler than initially anticipated.
Conclusion
This research began with a search for economic variables that could forecast Canadian consumer insolvencies. Household debt, debt-service costs, unemployment, inflation, income, wages and interest rates were all logical candidates.
More than 40 regression specifications and multiple time lags were examined.
That work was useful, but it ultimately led to a simpler and more interesting conclusion.
For forecasting Canadian consumer insolvencies one quarter ahead, the strongest predictor is the previous quarter’s insolvency rate. The relationship is highly persistent. Adding quarter-of-year seasonality materially improves forecasting performance.
The resulting production model currently explains approximately 93.4% of the historical variation in quarterly insolvency rates. More importantly, it has been tested through 117 rolling one-quarter-ahead forecasts. Its current historical Mean Absolute Error is 0.0411, RMSE is 0.0691, median absolute error is 0.0247, and mean signed error is close to zero at -0.0018. Adding quarter seasonality reduces MAE by approximately 22% compared with using the previous quarter’s rate alone. RMSE falls by approximately 9%.
The model has also now completed its first genuine live test.
It forecast 37,783 consumer insolvencies for Q2 2026. The actual result incorporated into the production dataset was 37,519—a difference of only 264 filings, or approximately 0.7%.
That result is encouraging but should not be overstated. One successful forecast proves very little by itself. Its real importance is that the model now has a live track record.
The second live forecast is 35,604 Canadian consumer insolvency filings for Q3 2026. When Q3 results become available, the forecast will be compared with the actual result and added to the record.
That process—forecast, observe, measure the error and test again—is ultimately more important than any individual regression statistic. The objective is not to find an equation that appears impressive when fitted to the past. It is to develop a forecasting process that is transparent, reproducible, falsifiable and capable of demonstrating over time whether it actually works.
That is now possible.
Methodology and Data Sources
Consumer insolvency statistics used in this analysis are derived from data published by the Office of the Superintendent of Bankruptcy Canada. Population and economic series used in the research are derived primarily from Statistics Canada.
Statistical analysis and forecasting are performed in R.
The production regression forecasts quarterly Canadian consumer insolvency filings per 1,000 population aged 15 and over using the previous quarter’s insolvency rate and quarter-of-year seasonality.
The predicted rate is converted into a filing count using population aged 15 and over for the forecast quarter. Where the forecast-quarter population observation is not yet available, population is estimated using recent quarterly population growth and identified as an estimate.
Monthly estimates are allocations of the quarterly forecast based on average monthly filing shares over the most recent ten-year period.
Historical forecasting performance is evaluated using an expanding-window procedure beginning with an initial 40-quarter training period. Each subsequent quarter is forecast using only preceding quarterly observations.
The model is intended as a statistical forecasting tool, not a causal model of consumer insolvency.
Use of Artificial Intelligence
Generative artificial intelligence was used to assist with programming, interpretation of statistical results, model development, and preparation and editing of this article.
The underlying data, statistical analysis, model-selection decisions and conclusions were reviewed by the author.