The first part of this series examined issues you need to consider before you approach deposit repricing modeling, e.g. what distinctions in your deposits base are important for your repricing. Those factors likely include differences in deposit types, such as interest checking negotiable orders of withdrawal (NOW) vs. money market deposit accounts (MMDAs), and may also include differences based on your depositor behavior or your competitors’ behaviors. The second part focused on the data required for a state-of-the-art repricing model: a history of your institution’s total balances, rates paid and balances retained in a given set of accounts. The third focused on the modeling process and the output from the estimated model.
The last part of this series focuses on monitoring, validating and updating the results. Core deposit repricing and deposit supply for most institutions has changed substantially over time as exhibits in the prior sections have suggested. It’s likely a major mistake to assume that the future does not hold further changes and challenges. Thus, institutions need to accompany their model development with a plan to actively monitor the model to maintain accuracy in capturing potential changes in repricing behavior and depositor behavior. A monitoring program should serve as an ongoing validation of the existing model or suggest changes needed to the model.
The starting point for checking the validity of a model should be backtesting. If a forecast is generated through 2018, its performance could be readily evaluated at the end of 2019. If during 2019 there was a +100 basis points (bps) change in market rates, then the 2018 forecast in the +100 bps scenario should be compared with the actual 2019 values. The two sets of data could be inspected graphically to give a visual perspective on the accuracy of the forecast. Actual and predicted values for rates and deposits should approximately match, and any adjustments to the rate shock should have approximately the same response, e.g. the lags and betas should be the same. The institution should identify and examine any discrepancies. A more sophisticated approach would be to adopt a statistical approach such as root mean squared errors (RMSE) to examine the differences between the actual and the forecasted values. RMSE calculates the square root of the sum of squared differences between forecast and actual values. When examining how a model performs over time, analysis of the RMSE from forecast to forecast will indicate whether the model remains on target. A stable model will yield similar RMSEs over time.
Backcasting also lends itself to policy corrections. For example, suppose your actual balances exceed forecasted balances. If your goal is to maintain constant balances, the backcast indicates that one or more deposit rates could be reduced to maintain balances. With rising rates, small discrepancies when identified early can lead to substantial cumulative returns.
An alternative to backcasting is retrocasting. While it is a better representation of the accuracy of the forecast, it is also more expensive. It compares actual values for rates and balances with forecasts from the model where those forecasts have been re-created using the actual values of all relevant variables. For example, if 2019 saw a +75 bps change in market rates that occurred gradually over the year, the Fed’s 2019 expectation at the end of 2018, a backcast would look at the 2018 +100 bps scenario and compare that with the actual values. But the two forecasts will differ not only due to limitations of the model but also because the inputs are slightly different, e.g. a gradual +75 bps change vs. a one-time +100 bps change. A retrocast solves the issue by re-forecasting 2019 but using as inputs the actual 2019 values observed for market rates. As with backcasting, statistics like RMSE can be calculated for retrocasts. Retrocasting also would potentially generate more accurate identification of discrepancies.
The final component of validation is updating. Ideally, the conditions under which a model is estimated will remain unchanged and the parameters of an estimated model can be used for forecasting for multiple years. In practice, however, the underlying structure on which a model is based is likely to change, sometimes substantially but sometimes in subtle ways. Using a model today that was estimated over the period 2000 to 2009 would clearly be problematic. Market rates are substantially lower; banking system excess reserves are substantially higher; and the presence of national banks in virtually all markets has increased substantially. But even a model estimated with data from 2010 through 2018 might change subtly in 2019 or 2020. Financial innovation, a larger footprint of national banks, a return to rapid growth or a recession could alter the model so carefully estimated at the end of 2018. Re-estimating the model – basically refreshing the parameters of the model – brings the model up to date. If there are any major changes that impact the model, that update should identify them and give you the opportunity to make any necessary changes. Best practices indicates an annual update is appropriate to maintain the accuracy of the model.
The quality of any analysis of core deposits for inclusion in an Asset Liability Management (ALM) model will depend on your commitment to accurately determine the behavior, the interest rate sensitivity, and the expected lives of your core deposits. A monthly history of deposit rates and deposit balances represents the absolute minimum data required. To more fully understand the behavior of deposits you should understand deposit retention and the factors that impact your retention rate. To do that requires a monthly history of retained balances. In addition, it is critical to understand the factors that distinguish the drivers of account balances, whether it is your rates or services, or direct deposits or automatic bill pay or competitors’ behaviors or Fed actions or some other factor. Those factors, combined with the final ingredient of situational awareness, will put you in a sound position to create a best practices model of your deposit behavior and to interpret and employ the results of that model to improve your bottom line.
Thank you for choosing the Situs Newswatch. If you want to see your company here or have an idea for coverage, please respond to this email or email email@example.com for more information.