The Globe and Mail’s top story on Wednesday suggested that CMHC is overvaluing the homes it uses as mortgage collateral.
It insinuated that the automated valuation model (AVM) built into CMHC’s “emili” underwriting system routinely overestimates property values. The implication is that taxpayers are at risk if mortgage defaults spike and CMHC cannot liquidate properties at their anticipated prices.
The story is portrayed like a scandal, which will likely undermine confidence in our housing market a bit more. Unfortunately, it’s yet another mortgage-related media story that is long on speculation and short on substance.
Before we begin, it’s worth noting that CMHC says emili (which has been around for 16 years) is not technically an AVM. Its main function is to assess overall borrower risk and not to determine a specific property value. We therefore use the term “AVM” loosely when referring to it.
AVMs exist for a reason and they’re used in dozens of countries. Their purpose is to generate objective and accurate valuations with less cost for the consumer, less managing of appraisers by lenders and much faster credit decisions. (CMHC often confirms property values in 7 seconds or less. That compares to 2-3+ days if a traditional appraisal is required.)
Now then, here is something that shouldn’t come as a shock: CMHC’s emili system, at times, overvalues properties. We all know that. It also undervalues properties (but that sort of thing doesn’t make for scintillating headlines).
emili’s job is not to pinpoint home values with 100% accuracy. No model can do that (nor can any appraiser for that matter). As such, over- or undervaluation alone is not the issue. What matters is the variance from true values. In other words, how much and how often is emili deviating from market value? If it’s 1% on average, that’s one thing. If it’s 10%, that’s another.
This data is unfortunately not publicly available. But you can be sure that regulators have it (or will soon). Anecdotally, we’ve heard that the percentage of properties valued more than 5-10% above actual appraised value is quite small, but there is no data to confirm it.
Of course, physical appraisals aren’t perfect either. On purchases, it is “extremely rare” for human appraisals to come in less than the purchase price, according U.S. research. (There are very few Canadian studies on this topic.)
On refinances, evidence suggests that appraisers may undervalue properties more than AVMs. That can reduce lender/insurer risk but it also has obvious downsides for refinancers (who may be unreasonably declined based on a bad valuation).
The tradeoffs between AVMs and human appraisals have been well-known for years. AVMs cannot easily evaluate factors like upkeep, view, property flaws, sun exposure, finish quality and other things that could add or detract from value. AVMs are most vulnerable in predicting values for remote, new or unique properties. That’s why insurers and lenders send out appraisers for properties that are more difficult to assess.
AVMs are also arguably less effective than appraisers for fraud prevention (assuming the appraiser is not involved in the fraud). The research differs on this point, however.
On the other hand, AVMs benefit from being emotionless machines. According to a U.S. National Appraisal Survey in 2007, over 90% of appraisers admit they’ve felt pressure to return a specific property value. By contrast, AVMs are completely uninfluenced by bias or pressures that afflict human appraisers.
While AVMs sometimes misjudge individual home values, there is ample evidence that they effectively value properties on a portfolio (i.e., overall) basis.
emili has evaluated millions of properties since its inception. CMHC says the system logic is based on:
CMHC calls its database “the most comprehensive…in Canada.” It includes property information on approximately eight million homes.
Despite what detractors believe, this isn’t a game for lenders and insurers. Erroneous property valuations are clearly linked to higher default losses. That’s why some lenders have reportedly increased usage of human appraisals to reduce their risk—even on insured applications.
Unfortunately, we have yet to find Canadian data that quantifies the difference between human appraisers and automated valuation systems (AVMs). There is plenty of research from the U.S., however.
Some studies have found that certain AVMs overvalue properties over specific timeframes. The median overvaluation in one study we saw was 4%. In addition, independent rating agencies like Fitch see it fit to deduct 5% from property values when those values are arrived at without full appraisals.
On the other hand, there are U.S. studies like this suggesting AVMs overvalue properties less than human appraisers—even in falling markets.
Either way, you can bet that insurers who rely on these models know the risks and adjust for them to the best of their ability. That may be why the ratio of auto-approvals at CMHC is down as of late.
Despite all of this, every lender executive we’ve talked with asserts unmitigated confidence in CMHC’s emili system. Indeed, some lenders we spoke with today questioned the very basis for the Globe’s story—which seems to be based on a few comments submitted last spring to OSFI. Those commentators were unnamed and their motivations are unknown.
Many also question the timing of this story since OSFI already addressed AVMs months ago in its B-20 guidelines.
In the coming days we’ll undoubtedly hear housing finance critics cite the Globe’s article as “proof” that CMHC’s valuation mechanism is reckless. But those charges are completely unsubstantiated based on the available evidence.
If proof ever materializes that CMHC is consistently and materially overvaluing properties, we’ll be among the first to report it. But we highly doubt that to be the case—in part because we routinely see more undervaluation than overvaluation (with our own clients).
We’ll be investigating this story further in coming weeks. But for now, it seems irresponsible to publicly discredit CMHC’s valuation model without clear data to back it up.
AVMs exist for a reason and they’re used in dozens of countries. Their purpose is to generate objective and accurate valuations with less cost for the consumer, less managing of appraisers by lenders and much faster credit decisions. (CMHC often confirms property values in 7 seconds or less. That compares to 2-3+ days if a traditional appraisal is required.)
Now then, here is something that shouldn’t come as a shock: CMHC’s emili system, at times, overvalues properties. We all know that. It also undervalues properties (but that sort of thing doesn’t make for scintillating headlines).
emili’s job is not to pinpoint home values with 100% accuracy. No model can do that (nor can any appraiser for that matter). As such, over- or undervaluation alone is not the issue. What matters is the variance from true values. In other words, how much and how often is emili deviating from market value? If it’s 1% on average, that’s one thing. If it’s 10%, that’s another.
This data is unfortunately not publicly available. But you can be sure that regulators have it (or will soon). Anecdotally, we’ve heard that the percentage of properties valued more than 5-10% above actual appraised value is quite small, but there is no data to confirm it.
Of course, physical appraisals aren’t perfect either. On purchases, it is “extremely rare” for human appraisals to come in less than the purchase price, according U.S. research. (There are very few Canadian studies on this topic.)
On refinances, evidence suggests that appraisers may undervalue properties more than AVMs. That can reduce lender/insurer risk but it also has obvious downsides for refinancers (who may be unreasonably declined based on a bad valuation).
The tradeoffs between AVMs and human appraisals have been well-known for years. AVMs cannot easily evaluate factors like upkeep, view, property flaws, sun exposure, finish quality and other things that could add or detract from value. AVMs are most vulnerable in predicting values for remote, new or unique properties. That’s why insurers and lenders send out appraisers for properties that are more difficult to assess.
AVMs are also arguably less effective than appraisers for fraud prevention (assuming the appraiser is not involved in the fraud). The research differs on this point, however.
On the other hand, AVMs benefit from being emotionless machines. According to a U.S. National Appraisal Survey in 2007, over 90% of appraisers admit they’ve felt pressure to return a specific property value. By contrast, AVMs are completely uninfluenced by bias or pressures that afflict human appraisers.
While AVMs sometimes misjudge individual home values, there is ample evidence that they effectively value properties on a portfolio (i.e., overall) basis.
emili has evaluated millions of properties since its inception. CMHC says the system logic is based on:
“physical characteristics of the property, the municipal property tax assessment, historical and current sales activity, and prior sales activity of the property being assessed, when available.”
“(emili) does not use property value averages, but uses the specific characteristics of the property being assessed. The database and models are continually updated and independently reviewed by a third party.”
CMHC calls its database “the most comprehensive…in Canada.” It includes property information on approximately eight million homes.
*******
With its reputation and profitability hanging in the balance, and with regulators keeping it under a microscope, we’re quite certain the country’s biggest mortgage insurer is not about to take shortcuts—not with hundreds of billions of dollars in real estate on the line. If anything, it will be overly conservative on values going forward, especially now that prices are softening.
Unfortunately, we have yet to find Canadian data that quantifies the difference between human appraisers and automated valuation systems (AVMs). There is plenty of research from the U.S., however.
Some studies have found that certain AVMs overvalue properties over specific timeframes. The median overvaluation in one study we saw was 4%. In addition, independent rating agencies like Fitch see it fit to deduct 5% from property values when those values are arrived at without full appraisals.
On the other hand, there are U.S. studies like this suggesting AVMs overvalue properties less than human appraisers—even in falling markets.
Either way, you can bet that insurers who rely on these models know the risks and adjust for them to the best of their ability. That may be why the ratio of auto-approvals at CMHC is down as of late.
Despite all of this, every lender executive we’ve talked with asserts unmitigated confidence in CMHC’s emili system. Indeed, some lenders we spoke with today questioned the very basis for the Globe’s story—which seems to be based on a few comments submitted last spring to OSFI. Those commentators were unnamed and their motivations are unknown.
Many also question the timing of this story since OSFI already addressed AVMs months ago in its B-20 guidelines.
In the coming days we’ll undoubtedly hear housing finance critics cite the Globe’s article as “proof” that CMHC’s valuation mechanism is reckless. But those charges are completely unsubstantiated based on the available evidence.
If proof ever materializes that CMHC is consistently and materially overvaluing properties, we’ll be among the first to report it. But we highly doubt that to be the case—in part because we routinely see more undervaluation than overvaluation (with our own clients).
We’ll be investigating this story further in coming weeks. But for now, it seems irresponsible to publicly discredit CMHC’s valuation model without clear data to back it up.
