Hi, I'm and welcome to my talk on machine learning management. I have over 13 years of experience in finance, digital transformation, focused on enterprise performance management tools and expense optimization. And today I want share some lessons I've learned the way in my career journey, and I hope this will be useful for most of your listening. In today's Asian world, we see many organizations focused on optimizing and reducing expense base as a way to improve their profit margins and give a better return to their shareholders. However, we see many financial teams such as F teams struggling with inaccurate and inefficient processing of expense data. They spend excessive time with reconciliations. While strategic decisions are hindered by incomplete and inaccurate expense forecasting, some of these challenges can be addressed by utilizing machine learning models to transform how businesses manage, analyze, and optimize their expenses. I've used financial services as a reference for many of my slides here, but these concepts apply to pretty much every industry out there. So without further ado, let's just. So to set a baseline, let's talk about some problems with the traditional expense forecasting methods where there is no utilization of historical data trends and patterns to refine your expense forecast. So with those traditional methods, we see about 27% of forecasting accuracy, which means your expense forecast deviate that much from your actions, undermining your budget process. We see about 42% classification. So these are pure expense categories, categorization, classification, and it could lead to significant misallocations of expenses across departments and their performance. And overall, this leads to high budget volatility, meaning the way when you start off your budget exercise and variation in. The budget exercise. So as you can see, traditional methods fall of expectations, and we need a smarter approach. So what is that smart approach? Let's just get there. Now. This approach is basically categorized into four different steps, and it's a sequential process here. Starting with organizing your raw financial data, which includes your expense data into structured taxonomies and hierarchies, which make it an optimal setup for machine learning analysis. And once you have your data ready to go, you can deploy any basic outta the box machine learning model, or you can build your own custom models to analyze that data and identify some hidden patterns, correlations, and an anomalies with any complex expense. And after your initial analysis, you can generate some insights and discoveries using some strategic business intelligence tools like dashboards or even simple analysis tools like Excel to give you some insights that help you make some strategic decisions which drive measurable cost optimization decision. Input make it. So let's just go deeper here. And before we go into side of things, I wanna upon a few definitions and way we define expense and department hierarchies in organization. Forecasting with machine learning works the best. So my first slide here is about cost and profit centers. And this concept I know is not very new. Every organization, every company out there has a concept of cost and profit centers, but I've seen a simple missing in many organizations, which I wanna highlight here. So I'll start off with the basic definition of cost centers. As you think of these are the departments or in terms of data, the fields where you book your expenses, and these are expenses that could be generated while you produce your revenue, or it could be overhead costs necessary for running a business, which includes the pure cost centers such as hr, IT facilities. And any nonpersonal expenses such as data fees or sub subscription fees. So in short, cost centers are where you book your expenses. And moving on to the next definition here, which is profit centers. These as a name suggest strategic business units where you. Front trading or sales divisions or product line divisions where you produce or make a sale and generate revenue. And that revenue is against these profit center quote. Now, the nuance I wanna highlight is every profit center should and must have a equivalent cost center where you book your expenses against. So that means every profit center has a cost center, but then not every cost center has a profit center. So those pure cost like it, they revenue need code for assuming that is clear, let's get to the next set of business definitions. Business definitions. So I wanna talk about the expense hierarchal framework here, which divides expenses into three different categories, starting with direct expenses, which make about 38% of any organization expenses. These are costs that tied specific to revenue activities with the, with clear ownership of where the revenues are coming from. So these could be personal compensation and benefit expenses related to people and teams in profit centers. Or it could be direct revenues produced in revenue units such as license costs or data, subscription fees and things like that. So in short, direct expenses are booked to. Gets allocated to now. The next bucket is allocated expenses. This sounds a little bit like debt expenses, but here I'm more focused on the organization overhead costs, which you can think of the technology costs, the it costs the compliance, legal risk teams cost, where these are all essential for a success. And these are peer cost centers like we spoke about in the earlier slide. And these could have some expenses, which could be confused as direct expenses because all these functions have dedicated personal. So all those expenses are direct expenses in those cost centers. But then at the end, all these expenses get allocated to frontend profit centers, which would then help you. Expenses are then waterfall method flow expenses, where when department allocate to another department, it only works one way and. So these overall expenses make about 45% of any organization's expenses, so roughly half. And that's very standard. And then the third bucket here is the variable expenses. So these, as the name suggest, are variable and they scale up as your organization or your company scales up and these are produced as and when you're producing revenue. It could be any non overhead costs that are directly a function of your revenues. So these exceptions are a, these expenses are a bit of an exception because these are produced during revenue generation process. So these get booked directly to profit centers. So this is a slide deviation from what we discussed earlier. I would. So now this is established. Let's move on to the more technical aspects of using machine learning models. So the first step for building any technological solution to, to make your business better is to have a strong data foundation. Is to identify your required types. So in this case, we're talking about forecasting. So we need have clearly defined transaction metta we need historical categorization between, it real estate and so on. Categor expenses have to. We also need vendor information. So these are third party vendors and companies that we make payments to. And once you identify those data types, you need to prepare your data to be able to be effectively used in a machine learning model. So that means you need to standardize your transaction formats handle missing values in your data. This is, in other words, mastering your categorization data. And we also need to normalize your vendor information in terms of formatting and so that the machine can recognize it. It's a vendor information every time you input it. And then once you have these steps done, you need to build your data structures around your transaction records, which unique such as primary keys and foreign keys. And you. Required for your organization. So if it's a financial services firm, you might need some external data as reference data. So that's an example there. So any robust machine learning implementation requires comprehensive data span at least 12 to 18 months before you can utilize that to do your budget reporting forecast. And as they all say. How properly your data is structured and how it, how your math data is defined, and how your data is clean and all, all well defined. So with that, once we set up your initial, your groundwork to set up your initial data structures. I wanna talk about a simple low hanging fruit use case of utilizing a quick machine learning model to make your job better as a financial analyst or an team member. So one such use case is your, is an automated class classification using a machine learning model. So this is simply helping you automate the way you categor expenses. And with those transactions, you clearly define your categories of expenses, spoke before between it, hr, your reimbursement, your travel costs, and so on. And then you engineer your features around those expenses. So this could be your transaction amounts, your vendor timing patterns, your payment methods, and also. And using another use case for. So this anomaly detection here. So these are simply identifying any outliers or anomalies in your expenses. So an example would be a spike in it to a product line unit. And this could be a allocation or could be a misallocation. So this model identifies that, and for that work you, you need to clearly define your features. So that would include involve transforming all your raw transac data into categories, your master metadata defined, and then for your model itself around how much your expenses are deviating from expectations. And you can also set up. Routing with expenses and have them reviewed by, and you could also give feedback them or help them better identify anomalies in the future. So that means reducing your false positives and making your machine work. Machine learning model work better with each cycle going forward. So typically, anomaly detection system employs multiple sophisticated approaches. They could be as simple as any unsupervised learning algorithms, or you could also use statistical techniques like which are applicable for any distribution functions. Further and deploy advanced machine learning or deep learning auto really complex patterns with extensive financial data. And as you could, as you can imagine, each model should be customed your organization size, complexity. Model doesn't fit for everybody is what the point is. And with those models you could also define dynamic threshold. So the model incorporates any seasonal patterns on both while customizing sensitivity controls to allow precise tuning by expense category to minimize false positives. So these are some really like high level use cases, which could be quickly achieved and established and make a big difference to your day to day expense management processes. Now let's talk about a bit more complex use case. So this is the time series forecasting up expenses, so where you would forecast your expenses. Where budgeting and forecasting into play. So as mentioned, this complex selection on timeframe, so starting with the smaller timeframes for about one to three months, which is a quarter. You could use models such as the Arima model, which is the integrated moving average model, which is more suitable for stable expenses such as it costs, which are more or less predictable and for the not so stable expenses that compliance costs, which are more variable depending on the business. Forecasting, which again, would take into account and those longer term, anything over one year or longer, you might have to use an ensemble of. For each of these models focused on individual timeframes, you could use metrics such as the MAPP for any accuracy measurement. And these are powerful metrics to help you identify if your model is working well. And also to give it the right feedback to. Defining it further as you go into your process and use them for dayday use cases. So again, I wanna highlight that for all these work you need to have your data those, you need to have your data recognized and. So with that I wanna also like recap on the implementation roadmap. So starting with the data assessment. So this goes this back to the step of organizing and figuring out your data sources. So your first step would be to assessment to. Think about buying a machine learning model. And once, once you have the step next, the next step would be to assemble and create a of engineers and specialists and financial analysts. And this step is really the most important step, and that's where many organizations fumble because it's very quick and easy to. But then having that team that uses that to its fullest extent and generates value the model, and assuming you have your team set up then the next step would be to. Deploy, deployment needs to be approach. Don't wanna build a complex machine, wanna gains and some champions across. So let's consider. Of the simple models with automated expense, categorization and detection, and for those full time series expense forecasting like we discuss in the previous slide, I would give it on two to three months, assuming that models work well and. Keep in mind is to have a team, a process for continuous monitoring and improvement of the models. Because even if your models are working very well, there's a need to refine, recalibrate, customize as your organization changes or especially if it susceptible to any as. I wanna quickly touch upon the technology stack options available for implementing these machine learning models in your organization. So if you're a small to medium size business, there are many open source solutions available in the market for data processing. Data to visualize it in a way that makes sense to the business audience. And for deployment. Like flak, Docker are small to medium businesses, but if you're a more bigger business, there are many other enterprise solutions available. And nowadays all the p and tools out there come with inbuilt machine learning modules that you can quickly deploy and get. Factor that have to think about those readily models even those clouds and. You think about your data integration architecture. So you build your data pipelines for ETL from data sources. You need to build a feature store for training your ML model, and then a model registry for deployment. And also think about any a layers that have to build to connect systems, so of finance, build those systems. Manufacturing, build a keys and your supply chain units tools. Make sure effectively, and again, like I before monitoring. So with that, I wanna close I some case. So from finance. Achieved remarkable improvements in classification. So it reduced its class misclassification rate from 30%, 6%. And in that process, an annual savings of million, just simply from better cost visibility and optimization that follows. They were able to achieve 76% reduction in processing time from just utilizing a huge manpower for categorizing and reconciling your expenses. So the machine model and freeing up the time for a team to focus on. This was all achieved within like an 18 month timeframe and three to five, and the initial values, the low three months. So I hope this inspires you and this was for you to understand and.