Limits of Machine Learning for Long-Horizon Deposit Forecasting
Summary
The document considers using recurrent neural networks and macroeconomic inputs to forecast bank non-maturity deposits over a multi-year horizon, including whether core deposits must be modeled separately. Its main caution is that financial time series are difficult to forecast because they can be nonstationary and noisy, while deposit history is too short to provide vast training samples. Machine learning can interpolate familiar conditions yet struggle when asked to extrapolate into unfamiliar regimes, so apparently adequate forecasts in stable periods may fail during major changes.
It recommends checking signal quality and overfitting, using time-aware training, validation, and test periods, and comparing the neural model with a conventional econometric baseline. Core deposits can be included or omitted and their explanatory contribution compared. Ensembles may also be considered, with interpretability as a tradeoff. The discussion is methodological advice, not evidence from a deposit forecasting experiment; it does not establish that neural networks will work or identify a universally best model.
Key ideas
- Financial time series can be noisy and nonstationary, limiting the reliability of learned patterns.
- Limited historical deposit data creates a risk of overfitting complex neural networks.
- Models that interpolate stable conditions may extrapolate poorly during unfamiliar regimes.
- Compare forecasts with an econometric baseline and use time-aware validation and testing.
- Test whether core deposit measures add explanatory power rather than assuming they are unnecessary.
Tags
Full text
# Forecasting non-maturity deposits with machine learning # Forecasting non-maturity deposits with machine learning I need to forecast non-maturity deposits in a bank. My intent is to use Recurrent Neural Networks (aka deep learning) to model time series. The model will learn from past bank data and macroeconomic variables, and forecast deposits in the next three years based on projected variables. Two questions: - Do you see a problem with this approach? - I believe that using this technique there's no need to calculate core deposits level to forecast total deposits, is this assumption correct? ## Answer by David Duarte (score 9, accepted) https://quant.stackexchange.com/a/60452 This boils down to timeseries forecasting which is particularly challenging for financial data because finance exhibits all the things that makes forecasting hard: non stationary, low signal to noise, etc. In the case of forecasting deposits, I very much doubt that you will be able to get satisfactory results even if you do manage to figure out all the relevant variables. And then you will have a problem called curse of dimensionality because you will never have millions or thousands (or even hundreds) of years of data to train your model. One thing to be aware about machine learning algorithms in general is that they are (can be) very good at interpolating and pretty bad at extrapolating, which means your model will fail when you need it the most. So my answer to your question "Do you see a problem with this approach?" is that you might get ok forecasts when is doesn't matter, ie, when deposits are not really changing, but forecasts may fail miserably when you are in unchartered territories. And this goes for forecasting just a few months in advance... never mind three years. Regarding the second question, I would try with and without core deposits to see if there is any explanatory power. A useful starting reference would be tensorflow's tutorial on Time series forecasting which uses Convolutional and Recurrent Neural Networks to forecast weather. You could try to use it with your data to forecast out of sample periods. https://www.tensorflow.org/tutorials/structured_data/time_series ## Answer by Sharad (score 4) https://quant.stackexchange.com/a/60430 Responding to the first question and building on Dimitri's comments, it is probably worth your while to do some due-diligence on the some of the standard issues that come up in Financial ML and see to what extent they apply to your problem: - Signal-to-noise ratio: Financial data sets are notorious for having low signal-to-noise ratios. Is your ML algorithm appropriately accounting for macro changes in industry structure, regulatory actions etc. - Over/under fitting: How do you control for time series effects in constructing your training, validation and testing data sets? Of course, these issues are not unique to ML models. You may find it useful to construct a baseline "adversary" for your model using some standard econometric approach and then compare and contrast how well it performs with respect to the above considerations (among others). You also don't have to get locked into an "either-or" approach, model ensembles often give better results that any single individual model although we get dragged back into issues of interpretability again. Although its specific focus is on asset pricing, you may find it useful to peruse de Lopez de Prado's book Advances in Financial Machine Learning which, apart from referencing these issues and discussing how to deal with them, also talks about how to structure a Financial ML project.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.