How to Model a Digital Twin of a Market through Agent Based Modelling
Maya Mateva19:582023
Test a market strategy in seconds, not the months a real campaign takes
Stop advertising on Facebook and it can take months to learn whether that helped, while you lose time and money finding out. Maya Mateva, co-founder and software engineer at Scios.ai, builds digital twins of markets: thousands of virtual consumers with their own drivers, perceptions and word of mouth. How are those agents built and trained well enough to guide launches for brands like Toyota and Netflix?
Learn how agent-based models turn segment data into a market you can test strategies against
Maya Mateva, co-founder and software engineer at Scios.ai, explains how to build a digital twin of a consumer market using agent-based modelling: thousands of virtual consumers built from segment data and trained on historical choices. She argues that such models let companies test pricing, launch, messaging and media strategies in seconds and with zero real-world risk, illustrating the approach with Toyota and Netflix case studies.
“With such platforms and such digital twins, you can test those hypotheses immediately, within seconds.”
“So thousands of diverse agents are created, and they're at the base of the market, this is this is the basis upon which you start creating your digital twin.”
What you will learn
Explain how agent-based modelling simulates individual consumers to reveal market-level outcomes
Identify the segment data needed to build virtual consumers: behaviors, drivers, perceptions, touchpoints and decision style
Recognize behavioral patterns such as defaults and the decoy effect that shape consumer choices
Plan zero-risk scenario tests for pricing, product launches, new competitors and advertising spend
Describe how decision intelligence combines behavioral economics, network science, attribution and machine learning
Key takeaways from the speaker
AI is a proven tool for efficiency in businesses demonstrated by real life use cases
2023 will be the year in which businesses will race to adopt AI Analytics to maximize efficiency and minimize costs
Digital twins can serve as a cost-efficient, lower-error alternative to external consultants
Details
Level
Beginner
Format
Talk
Technical depth
Introductory technical
For
Mid-career, People manager, Executive
Length
19:58
Published
June 2023
Questions this talk answers
What is agent-based modelling in a market context?
It is a simulation technique that models the behavior of single consumers as agents. Run across many agents, it reveals market-level effects and recreates the whole market, including competitors.
How are the virtual consumers in a market digital twin built?
Each segment's behaviors, drivers, perceptions, awareness, touchpoints and decision-making style are captured from representative data and assigned to agents. The agents are then trained and calibrated on historical data such as past choices and word of mouth.
What can a company do with a digital twin of its market?
It can test strategies with zero risk and within seconds, such as changing prices, facing a new competitor, launching a product without historical data, adjusting messaging or setting a target like 10% higher sales and letting algorithms search for the best plan.
How did Netflix use this kind of platform, according to the speaker?
Facing subscriber losses, Netflix evaluated different strategies and the software recommended 80 original series. The speaker said Netflix produced 72 and members came back.
What is decision intelligence?
It is a part of AI that helps organizations make better decisions, using techniques such as agent-based modelling and digital twins to create strategies that benefit both the company and its consumers.
About this talk
I have over 15 years of experience in engineering systems which deploy agent-based modeling for business decisions. I can take a deep dive into the mechanics of key considerations when building an AI-based software, as well as give practical examples of how we have augmented models in real life scenarios. I also envision telling the audience about a case study where we have helped a company form its business strategy using AI and digital twins like when we advised Toyota on the launch of its Mirai vehicle or when we helped Netflix form it's strategy to launch original content to address subscriber churn.
The behavioral shortcuts behind every purchase (59 s)
Test a strategy in seconds, not months (88 s)
How a simulation pointed Netflix to original series (60 s)
Maya Mateva: Oh, hi. I hope you can hear me. My name is Maya Mateva. I'm an engineer and co-founder at a company called SaaS.ai that develops a platform for decision intelligence for consumer market. And today I'm going to talk about AI, the basis of our platform that is digital to internet. I hope you can see it, but yeah, I'm going to share with you how to create the digital twin of a market through agent based modeling.
So imagine that you wake up in the morning and you have a million, million decisions to make. You have a million, million decisions to make. What to eat, where to go to, how to travel, where you should buy your lunch for, whether you want a Pepsi or Coca-Cola or not. And as we, we, we are met with so many decisions, we developed some behavioral patterns in order to help us beat the decisions, let's say.
For example, we do prefer defaults or the decision of not making any decision. Meaning that if we all have the checkbox, do you want that email preselected, we will probably end up with a lot of spam in our folders. We shape our perceptions based on expectations, whether we think something is beautiful or not, soft or not, and so on.
We tend to think that our, the things we own are with much greater value than those, those that we not, because we spend some effort and affections to them. And we create our own relative reality, determining the value of something based on something else. And then we create our own approaches based on our behavioral patterns. For example, they use our formal for making us want the latest iPhone, for example, or they influence our, our defaults or not just to do something by designing environment that makes some decisions easier than others.
And then they introduce additional options, making us choose something for example, Starbucks, something that I drink every day, but they use the so-called decoy effect, and they introduced the medium, the medium beverages. They're small medium and big, the price between medium and big is pretty much the same, but you know you get more when you buy, when you buy the bigger one.
So you naturally choose the bigger one because it's, it's the offer right now. So companies simplify the messages that they, the text, the key information that we need to absorb. They suggest selling this suggest and encourage us to make a decision in the spur of the moment, so that we just do it and buy something. But even though our behavior patterns are pretty well known, and there are a lot of books like predictably irrational.
Nudge thinking fast and so all those books about the way we behave does not make companies making a winning strategy easier. Let's say the thing is that the world right now. It's pretty complex. There is a lot of data and companies need to absorb that data and act upon it. But right now, actually a couple for a while now, there is, there is technology, there is AI, something that can help us make decisions based on that data.
And with consumer markets and with consumers, the thing, the thing is agent based modeling or creating real world consumer simulating our behavior. So introducing agent based modeling it's a simulation technique that creates micro consumer behavior through agents don't cover macro marketing outcome, meaning that it simulates the behavior of single person. And when that's done for a lot of person for a lot of individuals, we can get the micro effects.
So how it goes recreating consumers requires basic information about the different segments. In a market, there is a population that population is divided into different segments, whether it's based on religion, skin color, age, and so on. We do tends to, to be some segment. And we have some typical behavior behaviors. We use something in, in a way, or we do do researches online or not.
We have our own profiles in that segments. What are our drivers? Whether we are aware of something or not? What are the perceptions for something? How we communicate? If we are between 20 and 30, we'll probably communicate more on TikTok and absorb advertisements more there. If we are, like me, between 30 and 40, you're keen on Facebook and Instagram and so on.
We put information about the touch points, all the places you see that ads and that information, whether it's about a brand, a prime minister, for example, or a TV show. And at the end, we can evaluate what are decision-making processes, how rational we are, whether we choose something after reading a lot about it or we choose something at the spur of the moment, like Coca-Cola, for example, whether we are emotional for our choices or not, it all depends on the target.
And a representative data set is collected for each agent so that you know it, what segments, what are the perceptions for those between 20 and 30, for example, what's the behavior for those between 30 and 40, for instance? What are our drivers, whether for some the price of something is what's more valuable and what we consider the most or rather it's the quality of something.
And we assign values to the agents for all those characteristics and for our drivers and perceptions. And that's how we create a consumer segment. And we assign values to different segments with different perceptions, different behaviors, different drivers. And we do that still we have representative populations of agents, needless to say, as much data as you have, the better, it's the best, because you have very accurate models and very accurate, accurate representations of the agents of these segments in your market.
So thousands of diverse agents are created, and they're at the base of the market, this is this is the basis upon which you start creating your digital twin. And those agents are trained based on historical data, of course, like every model, we we learn them, we train them, calibrate it, what are the perceptions, what are their past choices?
What are the drivers behind each segments? How they choose an alternative, how they experience it, do they talk about it or not? Agents segments with different ages will concentrate more talking online, while those advanced age, let's say we will probably talk with each other, and that's how they will share their experiences and talk. All those dynamics can be simulated through agent based models.
And the science behind search models, it's pretty amazing. Initially, it's the one I talked about, it's agent based modeling, it recreates the entire market, not just the performance of a single brand. So in search, in search ecosystem in an agent based model, you can evaluate not only, not only Toyota, for example, but Volvo, Mercedes and BMW, the same way it's the real world.
So with agent based modeling, you can create real world scenarios. There is again, the behavioral economy, economics that models consumer decision making processes at individual levels. Network science that accounts for words of mouth and social influences, attribution and analytics that considers all touchpoints simultaneously, not only watching Instagram or Facebook, but watching TV, listening to radio. All those dynamics can be can be gathered in a single model and at the end it's machine learning that calibrate in market results.
Through different algorithms so in its training with each, each new data that is provided to the model, so that the agent is as the virtual consumers are as accurate as possible. And when we have the agents, we are pretty good to go to create our digital, digital twin. We simulate the consumer decision process, how people, how people react, how they make decisions, what they see, whether they, they have, whether the distribution is sold that they can choose between.
Subarro and Subarro and Subarro and Subarro and Subarro and that's the best car, I guess, we can simulate how they see something where they listen to where they get information, no matter if it's for a car, a TV show, the new president or. Subarro and Subarro and Subarro and Subarro and Subarro and Subarro and Subarro. digital twin of the market through agent-based modeling.
The value behind that, once you have the market with its environment, with its alternatives, with the competitors, no matter if it's for a car, presidents, TV shows, or actor and actresses for a particular movie, for example, then we create, as I mentioned, virtual consumers with their perceptions, consideration, decision drivers. We create them as close as possible to the real people, to the real segments.
And once that's done in a platform, in a software, we can create meaningful strategies that leads to growth, efficiency, and optimization. That strategy that can tell us which activities to focus on. What is the most appropriate messaging? If we want to get a bigger market share, what we should do? If we want to launch a new energy, for example, like it was with Coca-Cola a couple of years ago, how to design it, what to do with it, how to position it contrary to Red Bull, for example.
How to distribute your things, your brand, as much as possible, as possible so that you get the best strategy for it. So, but what is more with such models, you can test strategies with zero risk. For example, if you stop advertising in Facebook, it's going to take pretty much time to see what's the result, whether that's beneficial or not.
But you will lose time and you will probably use money. Hopefully, that strategy may be successful, but maybe not. But it will take months for you to understand. With such platforms and such digital twins, you can test those hypotheses immediately, within seconds. You can test what will happen if there is a new competitor on the market. What will happen if you change the prices, whether it's going to be successful on the short-term or long-term as well, how this will affect your brand awareness, how to release and to launch a new product where there is no historical data about that before, how to improve your messages if you want to target different segments, how to modify your investment, because with such platforms, with platforms like ours, you can say, okay, I want to reach a certain target.
I want my sales to be 10% higher, tell me what to do, and the algorithms below will test different hypotheses, different scenarios, turn some activities on and off to see what is the best possible strategy strategy and suggest you such such optimal way to go forward. You can see and adapt more easily in platform events like COVID, COVID pandemic or worse, some unexpected, some unexpected things that may happen to us.
Or if you have had such a platform for decision intelligence, you can see what will happen if you remove the blue ticker, ticker, Twitter, tick or make it paid and how, how your consumers, how your users will react based on that. So the scenarios, as I mentioned, just a second ago, is they are endless. For example, in the COVID, we worked with a company that wanted to see whether people will change their behavior to drinking their patterns during COVID.
It turns out people will still go on a drink, but their patterns will behave, will change a little bit, and the bars, wine, wine shops, and so on, could adapt more easily with such information in advance, staying ahead of the curve, rather than waiting to see whether they will survive or not. You can test, you can apply such, such platforms in the case of voting, evaluating different presidents, for example, or wanting a new, a new product, or saying what will happen if you, if you change the price and make your, make your car, for example, more exquisite, more exclusive than others.
I'm going to delve a little bit more into two scenarios that are pretty much different than each other, but they all drill down to customers' decisions. The first one we worked on was developing a new product with Toyota. They developed a pretty accurate model, they had a lot of data, and they were about to launch a new electric vehicle model on the market.
But they needed to be prepared to know what features they should focus on, whether it is the aggressiveness of other cars or safety, what are the, what, what they need to focus on, that is most impactful and meaningful for their market. The second one, they needed to focus on, how they validated different pricing strategies, different messages, they evaluated competitor's actions as well, and they needed to prepare for Tesla coming, Tesla 3 being released at that time, so they, they needed to be prepared for that as well in order to be, to be the competition, and their model was pretty accurate.
The second one, they needed to be prepared for that as well. They needed to be prepared for that as well. The other use case of ours was with Netflix. Back in the day, they were pretty much about to sink. They lost hundreds of members and their goodwill. People were starting to started to stop watching Netflix. And at that time, their CEO, Reed Hastings, was facing pretty, pretty steep targets.
And what he did then was to use our platform to evaluate different strategies. The successful one was surprised that they should have original series. And they did. Our software told them that they needed to have 80 original series. They had 72 and that was pretty good, pretty great success. Their stops were high again. users, their members were coming back again.
And it all depended that because the people evaluated different, that that is something new. But with Netflix, they managed to test different, a lot of scenarios to see what is the more, the more prominent, the more optimal one, not only for short term sales, but for long term brand metrics like perception, consideration and word of mouth. And yeah, and they, they managed to stay and be alive.
To stay and be alive. All those platforms for all those scenarios that I mentioned, they are the. They are examples of the era of decision intelligence. Decision intelligence is a sub part of AI of artificial intelligence that help us make better decisions at the company and the organization level. In those today's complex and deserve that business environments using KI and as, as I just mentioned, agent based modeling digital to install those techniques and technology, you can create strategy that's beneficial for it for your company and for your consumers.
And nowadays companies are getting track of it. They, they, they got the experience. Pretty, pretty huge 60% increasing desired outcomes. AI simulation adoption is pretty amazing. 96% of the last companies to make it according to McKinsey plan to use simulations for forecasting market conditions. Make long term strategic decisions, develop new products and support financial sales and marketing plannings.
And what is more their confidence in the accuracy of their decisions and the path forward is rising and they feel pretty confident in what they do right now. So they will stay ahead of the curve and win as much market share as possible or depending on their strategies. So thank you and embrace the decision intelligence era.
Transcribed automatically; names checked against the speaker record. Select a time to jump there.
Article based on this talk
How a digital twin of your market lets you test strategy before you spend
Maya Mateva explains how thousands of agent-based virtual consumers can replace months of trial and error with simulations that run in seconds.
In short
Maya Mateva argues that companies no longer need to wait months to learn whether a marketing, pricing or product decision works. By building a digital twin of a market from thousands of agent-based virtual consumers, trained on segment data and historical choices, teams can test strategies in seconds and with zero real-world risk. She grounds the approach in consumer psychology, walks through how agents are created and calibrated, and illustrates it with cases involving Toyota and Netflix.
Key takeaways
AI is a proven tool for efficiency in businesses, as demonstrated by real-life use cases.
2023 will be the year in which businesses race to adopt AI analytics to maximize efficiency and minimize costs.
Digital twins can serve as a cost-efficient, lower-error alternative to external consultants.
Predictable consumers, unpredictable strategy
Mateva opens with a familiar experience: waking up to countless small decisions, from what to eat to whether to choose Pepsi or Coca-Cola. To cope, she explains, people develop behavioral shortcuts. They shape perceptions from expectations, overvalue what they already own, and judge value relative to something else.
"For example, we do prefer defaults or the decision of not making any decision." (1:09)
Companies, in turn, design around these patterns. Mateva points to the decoy effect in coffee sizes, where a medium option priced close to the large one steers customers upward: "So you naturally choose the bigger one because it's, it's the offer right now." (2:48) She notes that books such as Predictably Irrational, Nudge and Thinking, Fast and Slow describe this behavior well, yet knowing the patterns does not make building a winning strategy easier. The world is complex and data-heavy, and that is where she positions AI, and specifically agent-based modelling, as the tool that helps companies act on data.
Building a market out of virtual consumers
Agent-based modelling, as Mateva describes it, simulates the behavior of a single person, and when that is repeated across many individuals it reveals market-level effects. The work starts with segments. For each one, the team captures typical behaviors, drivers, awareness, perceptions and communication channels: younger consumers may absorb advertising on TikTok, while people in their thirties lean toward Facebook and Instagram. Touchpoints, from social media to TV and radio, are mapped, and so is decision style: whether people research carefully or buy on impulse, rationally or emotionally.
Values for these characteristics are assigned to agents until the population is representative, and more data means more accurate agents. "So thousands of diverse agents are created, and they're at the base of the market, this is this is the basis upon which you start creating your digital twin." (7:07)
Those agents are then trained and calibrated on historical data: past choices, how they experience products, and whether they talk about them online or face to face. "All those dynamics can be simulated through agent based models." (8:21) Mateva stresses that the model recreates the whole market, so a car simulation includes Volvo, Mercedes and BMW alongside Toyota. She lists the disciplines combined underneath: behavioral economics for individual decisions, network science for word of mouth and social influence, attribution analytics across all touchpoints at once, and machine learning that keeps calibrating the model as new data arrives.
Testing strategy with zero risk
Once the market, its alternatives and its virtual consumers exist in software, Mateva says, the twin can point to meaningful strategies: which activities to focus on, which messages work, how to position a new product against an incumbent, and how to distribute it.
"So, but what is more with such models, you can test strategies with zero risk." (11:45)
Her example is pulling advertising from Facebook. In the real world it takes months to see whether that helped, while time and money are spent. "With such platforms and such digital twins, you can test those hypotheses immediately, within seconds." (12:14) Teams can simulate a new competitor, a price change and its short and long-term effects, a launch with no historical data, or new messaging for a different segment. Platforms like hers can also run in reverse: set a goal such as 10% higher sales, and algorithms switch activities on and off to search for the best strategy. She adds scenarios from COVID, where a client explored whether drinking patterns would shift so that bars and wine shops could adapt in advance, as well as voting and premium pricing.
What the Toyota and Netflix cases show
Mateva describes two contrasting cases that both come down to consumer decisions. "The first one we worked on was developing a new product with Toyota." (15:09) Toyota was preparing to launch an electric vehicle and needed to know which features mattered most to its market, how to validate pricing strategies and messages, and how to prepare for competitors, including the release of Tesla's Model 3.
The second involved Netflix at a time when, as she tells it, members were leaving and CEO Reed Hastings faced steep targets. Netflix used the platform to evaluate strategies, and original content emerged as the winner. "Our software told them that they needed to have 80 original series. They had 72 and that was pretty good, pretty great success." (17:12) She emphasizes that the scenarios were judged not only on short-term sales but on long-term brand metrics such as perception, consideration and word of mouth. These results are presented as the speaker described them.
Entering the decision intelligence era
Mateva closes by naming the broader field. "Decision intelligence is a sub part of AI of artificial intelligence that help us make better decisions at the company and the organization level." (18:08) As the speaker cited, companies using these approaches report around a 60% increase in desired outcomes, and according to McKinsey, 96% of companies plan to use simulations to forecast market conditions, make long-term strategic decisions, develop new products and support financial, sales and marketing planning. Her closing message is simple: embrace the decision intelligence era.
Watch the talk
How to Model a Digital Twin of a Market through Agent Based Modelling, by Maya Mateva (20 minutes): https://www.womentech.net/video/how-model-digital-twin-market-through-agent-based-modelling
About the speaker
Maya Mateva is Co-founder and Software Engineer at Scios.ai, having previously worked at the US-based decision intelligence company Concentric, where she was vital to developing the platform Scios now deploys. She leads platform expansion at Scios, including the Goal Optimizer and a Storytelling feature, while pursuing a PhD in Enhanced Software for Aiding Disabilities at the Technical University of Sofia.
Questions this talk answers
What is agent-based modelling in a market context?
How are the virtual consumers in a market digital twin built?
What can a company do with a digital twin of its market?
How did Netflix use this kind of platform, according to the speaker?
What is decision intelligence?
Drafted with AI from the talk transcript. Editor review pending.
Frameworks and ideas mentioned
Agent-based modelling
Digital twin
Decision intelligence
Decoy effect
Behavioral economics
Network science
Attribution analytics
Resources mentioned
Predictably Irrational (book)
Nudge (book)
Thinking, Fast and Slow (book)
Cite this talk
Maya Mateva. "How to Model a Digital Twin of a Market through Agent Based Modelling." WomenTech Network, 2023. https://www.womentech.net/video/how-model-digital-twin-market-through-agent-based-modelling