A turning point in the debate on risk and return occurred in the early 1990s, when Gene Fama and Ken French wrote one of the most-quoted and influential papers on the topic. In it, they began with a simple premise. If our objective in risk and return models is to come up with expected returns on investments, should we not judge the quality of these models by looking at how well they explained actual returns over very long time periods? They began by looking at the CAPM: in it, all return differences across investments should be explained by differences in betas. Looking at actual stock returns from 1962 to 1990, Fama and French found that betas cannot explain a very large portion of the differences in returns across stocks.
This finding was not knew and reflected what other researchers had concluded in earlier papers. Fama and French, however, decided to turn the problem on its head. Rather than build an alternative risk and return model, which is what others had tried to do with the arbitrage pricing and multi factor models, with all their baggage, they decided to start with the data on returns and work backwards. In other words, they went looking for other company characteristics that would do a better job of explaining differences in returns across stocks, than betas did. Their search led them to two variables - the market capitalization of the company and it's price to book ratio- that together explained a large portion of the differences in returns. Small market cap companies and low price to book value companies consistently earned higher returns than large market cap companies and high price to book value companies. Rather than view this as an inefficiency (which other papers had in the past), Fama and French considered these variables as proxies (stand ins) for risk. In effect, they concluded that small companies must be riskier than large companies and low price to book companies must be riskier than high price to book companies.
While the logic of the Fama-French approach is impeccable, it has one significant weakness. Since it is data driven, any proxy, no matter how outlandish, that explains differences in returns could be used in the model. At the risk of pushing this argument to absurd limits, if companies with fatter CEOs have higher returns than companies with thinner CEOs, the weight of the CEO should be used as a risk proxy. Not surprisingly, as the data we look at gets richer and deeper, other variables have been added to the list of good risk proxies - return momentum, earnings revisions, insider buying etc.
Models that test the CAPM against proxy models by looking at how much of past returns are explained by each are almost certainly going to find the CAPM wanting. After all, the proxies were not picked at random but because they had been correlated with returns in the past.
I think that the question of whether to stick with the CAPM or go with a proxy model depends in large part on what you are using the model for. If your job is performance evaluation, say of mutual funds, I think that proxy models make sense, because you are looking at the past and examining whether individual mutual funds beat the market. If you are trying to forecast expected returns in the future, which is what you are doing when estimating cost of equity in corporate finance or valuation, I prefer to stick with the CAPM, with bottom up betas and adjustments for financial and operating leverage.