How we get one of the busiest ATM networks in the world to run like clockwork
Find out what happens behind the scenes in ATM operations and why you can always get cash when you need it
Cash is an integral part of our day-to-day life that enables us to carry out our daily transactions. The feel of a physical note appeals to our emotional psyche and gives us a sense of ownership. At DBS, we are the invisible partners for our customers in this journey, enabling them to “Live more, Bank less”. In Singapore, DBS runs one of the busiest ATM networks in the world, with 17 million transactions per month. That works out to about six transactions every second, or about 40 transactions in the time it took for you to read the previous sentence. According to the ATM Industry Association’s (ATMIA) global benchmarking study in 2018, DBS’ machine utilisation rate is four times greater than the global average. Though Covid-19 has resulted in seismic shifts in consumer behaviour and increased digital adoption, cash continues to dominate, and DBS’ machine utilisation rates are currently at 80% of pre-pandemic levels.
As we ride on the wave of digital transformation, transforming into a technology company that does banking, how can we use data and analytics to load our machines in a timely manner, ensuring that our 4 million active customers have ready access to cash?
Introduction to Cash Operations
In Singapore, DBS supports more than 1,400 customer touchpoints for withdrawing and depositing cash in partnership with our service providers, machine vendors and regulators. For a machine to be loaded with cash, we depend on our Cash Processing Centre (CPC) to prepare the cassettes with notes and our Cash-In-Transit (CIT) security guards to load our ATMs with the cassettes. We also manage the cash demand during major events, such as Chinese New Year (CNY), Formula 1, Covid-19 circuit breaker, etc.
DBS has a 24/7 Operations Management Centre (OMC) that works closely with our CPC and CIT teams. But 24/7 monitoring alone doesn’t allow us to improve our customer experience. Machine unavailability is the biggest customer pain point. This is most likely because the machine runs out of cash (cash-out) or does not have the preferred note denomination. To address these issues, we looked at the challenges to be managed and resolved:
1. Can we load all machines every day? If not, can we load on alternate days?
2. What if high-usage machines run out of cash?
3. What if some machines are not used at all?
We focused on our efforts on building a data analytics and machine-learning engine that leverages past historical data to forecast potential customer withdrawal / deposit demand and proactively load our machines in advance before they cash-out. This also enables our CIT and CPC partners to plan their resources accordingly.
How We Transform with Data
To minimise cash-out in our machines, reduce machine downtime and minimise customer impact, the cash planning model (CIT model) generates a replenishment schedule for our vendors that can accurately prioritise the machines to be loaded and the timeslots for when these machines should be loaded.
Architecture
The CIT model consists of various components, namely data ingestion, compute, data science module, optimisation, and user visualisation.
Data Ingestion
To know whether a machine requires cash loading, we start with a snapshot of the latest cash positions from all the DBS machines in the network. This data is updated onto a host Auto Banking System, which further sends the updated logs to our in-house data platform. During the ingestion process in the data platform, this data is transformed from a machine language to a more human-readable format with adequate metadata. A total of 28 data feeds were created, which are ingested daily along with 12 user input parameters act.
The 12 user input files are parametric variables for the data science forecasting models and the optimisation of these models. These files provide users with the capability to override the default configuration. Though the default configurations work on most occasions, these user input overrides help bring human intelligence and experience into the loop. It enables users to adjust the parameters and to cater to special occasions or events on the ground.
Compute Transformations
Once the source files are ingested into the data platform, the in-house computing and transformation engine then performs the Extract, Transform and Load (ETL) processes. Based on the metadata, the ETL engine transforms these data sets into granular datasets by creating data partitions and views at the machine level, hourly level, and denomination level. The data generated is suitable for data science modelling and user consumption.
Forecasting
As a precursor to forecasting, a few additional pre-processing checks such as null handling, missing value checks and scaling are completed. For null values, the datasets are imputed with median values from the recent month being used as a reference. A seasonal ratio of the past two years’ historical data is also computed as a scaling factor to prevent erroneous results or out-of-bound ranges during calculations.
The historical withdrawal and deposit data on each machine over the past years is used by the forecasting models to generate expected forecast values for the subsequent days. DBS machines provide both dispense and deposit capabilities and hence separate models are created for each machine type.
To forecast the demand for ATMs, we use a time-series forecasting model Seasonal Autoregressive Integrated Moving Average with eXogenous regressors model (SARIMAX) to forecast. Unlike normal moving averages (MA) or Autoregressive moving average which identifies a global minima using Ordinary least square method, SARIMAX uses Mean Absolute percentage Error (MAPE) to identify the error rate. For this business problem, SARIMAX addresses the seasonal variations across the year and weekday/weekend variations better when compared to other time series models such as LSTM, auto-ARIMA.
Operationally, when we moved the models to live production, we encountered some challenges. One, handling the cold start problem for newer machines where no historical data is present. This was resolved by using similar machines with similar withdrawal patterns or in the similar locality as a proxy for the new machine. Two, periodic model risk assessment and quality management have to be done to overcome bias in the models. Based on periodic model retraining and result validation, some model parameters are adjusted to choose the best possible model.
Improved Customer Experience
Timely replenishment of machines has helped to significantly reduce machine downtime (machine availability to our customers has been above 99% consistently for the past five years) and helped us prevent cash-outs. It has also encouraged us to understand customer usage, behaviour, and journeys better. Based on user feedback, we have also conducted additional experiments related to queue markers, signage for our machines and the optimum mix of withdrawal and deposit machines in a location. Following these experiments, we have implemented the relevant changes in the field, ensuring that our customers’ feedback is addressed.
Key Learnings on our Transformation Journey
Simple models are always better and easier to interpret
With abundant infrastructure and off-the-shelf model techniques available, product owners are more tempted to look at complex models in their endeavour to improve accuracy. Best fit models don’t necessarily give the best outcomes and account for all variations. Depending on the use case to be modelled, product owners and business teams can decide the complexity, nature of models and operational risk involved. In our case, by Occam’s Razor, we have chosen simpler models for easy explanation.
Costs vs Benefit trade-off
Any data science model is built on the cost vs benefit trade-off. The cost related to the choice of model, infrastructure used, and resiliency frameworks only make sense when it is associated with tangible business benefits. In our DBS story of digital transformation and data adoption, CIT forecasting model has helped massively reduced our operating costs, thereby enabling lower carbon footprint.
Keeping human intelligence in the forecasting model
The 12 manual override files ingested by internal user for forecasting and optimisation allows human intelligence and experiences to adjust some default parameters or override parameters to account for changes in the field. This has helped us overcome withdrawals during CNY periods, and one-off occasions such as Covid-19 circuit breaker effectively.
