The model will never be ready. Rather we build a new version practically every week. New products, research and companies emerge every day, and that is pretty cool. We are happy that already now, in our project’s infant phase, we are able to produce information and understanding that to our knowledge no other system can. However, the road ahead of us keeps us humble and excited.
The model will never be perfect. All models and numerical representations are just proxies, they ARE not the truth. A perfect model accurately depicting all impacts of all companies in real-time is theoretically impossible. However, a net value creation measuring system significantly better than the one at use today CAN be done. Not being able to build something perfect is a lousy excuse for not doing better.
We want to put together all the work of millions of researchers by using scientific articles as our main input data. Also, we want to build a system where everyone who wants to contribute to the building of better understanding of what companies achieve can do so. Our aim is to help people collaborate in making more fact-based decisions.Help teach our AI ⟶
This is our dearest design principle throughout the model. It means we only concentrate on the largest impacts a company has on the surrounding world. It’s the only way to stop losing sight of the Tier 1 priorities (e.g. fighting climate change) in Tier 2-10 data (e.g. whether or not your company serves cool artisanal coffee for its employees). Also, it’s a great way to stay sane. Practical example: for an oil company, we don’t care whether or not they use recycled office paper.
Analyzing and optimizing large groups of companies (e.g. investment portfolios).
Making comparisons between products or companies.
Understanding the scale of impacts.
Dividing companies into good and bad ones.
Comparing two brands of the same product with largely similar impact.
The Upright model aims to build a big picture of what kind of value companies create. Our approach can be simplified to being built on three inputs: an understanding of what kind of products and services exists, a list of all the ways a company can impact the world around it, and scientific journal articles as source data for understanding about relationships between the two.
Now comes the tricky part. We need data about the impacts that the products and services have on the environment, health of people, society and creation and distribution of knowledge. This is where we need our engineering skills: to teach a neural network to understand causalities in natural language, i.e., the scientific articles.
Over the last 6 months, we have built a pretty exceptional training data set of over 30 000 scientific articles which we have read and classified manually to train the neural network.
After getting the raw data from the neural network, we form scores for each combination of product and impact using an algorithm that combines information about 1) the prediction distribution, 2) number of articles studying a particular product/impact pair, 3) position and relative relevance of the product in its value chains, and 4) position and relative relevance of the product in its product family hierarchies.
After this, we have the basic building blocks for forming net impact profiles for all products and services. By summing these up, we get companies - which we can further sum up to form portfolios, funds, industries and other entities whose impact we wish to understand.
And that’s it! It’s a pretty crazy exercise, but we really believe this is problem worth fighting for. This is our first effort towards solving it. We invite you to follow our progress, cheer us on, contribute and/or build your own solutions!
When answering questions about companies’ impact on the surrounding world, one of the big questions is: where to get the data? Most organizations offering solutions today are using data reported by companies themselves. This is a pragmatic approach, as there is a lot of company data available.
Upright, however, is taking a different approach and building the backbone of its model based on scientific data. We want to make the role of marketing and branding communications smaller, and bring scientific knowledge to center stage. We also want to help facilitate a dialogue between the producers and practitioners of data - researchers and business leaders.
The first version of the Upright model forms the “backbone” for impact scores using data from three primary sources:
The Upright methodology of reading causalities in natural language and combining that information with direct numerical input allows us to gradually add many different types of data sources (e.g. other academic corpuses, news data). More about the status of the model today and what plans we have for additional data sources in the future can be read here.
One of the biggest challenges in measuring net impact is: how can we put all the different impacts into the same unit? How can we compare greenhouse gases, diseases and taxes to one another? Upright’s approach to this question is the following:
We start by building a model where all impact categories, such as GHG emissions, diseases or knowledge creation are assumed to be of equal weight and importance. This means that all 19 impact categories impact the final net score equally much. We call this representation “equal weights value set”.
Next we define the score for company X in impact Y to be: impact Y caused by company X / total amount of impact Y by all companies globally. In this way, we end up with (teeny tiny) quotients.
It is worth noting that this is an abstract concept. This is the ideal for which we seek proxies.
We then make it possible for the user to set their own optimization criteria. This means allocating 100 % of weight to all the impact categories in a way that describes what they want to achieve with their consuming, investing, working or business.
For more, see how the scores are formed.
Practically all solutions do some kind of comparing of different impacts and fitting them into the same unit of measure. However, many of them are using fixed value sets and not making the assumptions transparent to the user of the data. For example, ESG ratings produced and sold for investors form one rating per company describing how “sustainable” it is. The rating calculation is based on a set of assumptions: what is considered to be “impact" (typical answer: “environment, social and governance” factors), how important each of them is compared to the other (typical answer: “equally important” or other weights chosen by analysts) and what are used as proxies for whether or not the company is serving these impacts (typical answer: "whether or not they are complying to certain rules and standards"). Upright's aim is to make it possible for users to become aware of the values they are currently practicing - and to consciously strive for serving the values they actually want to serve.
In the net impact profile for a product or company, you can see numbers. These are what we call impact scores. There are two phases in forming them. In the first one, we define the idea of relative and absolute impact scores and what questions they aim to answer. In the second one, we try to find proxies to populate the perfect idea with imperfect data.
There are two types of impact scores: absolute and relative ones. Absolute scores aim to tell you about the absolute impact a company has on the surrounding world. Naturally, this number would be much higher for a large company than for a small one. That's why we also need relative scores. They tell you how much “bang for the buck” as, for example, an investor you get when investing a certain sum of money.
Example: Let’s imagine two almost identical apple pie bakeries. They both do things almost exactly the same and bake exactly similar delicious apple pies with exactly the same nutritional facts at exactly the same price using exactly the same suppliers and selling to exactly the same customers. But one of them is small with only USD 0.1 million of revenue. The other one is big, with USD 100 million of revenue. The relative score for these two companies would be the same, but the absolute would be 1000 times bigger for the larger one.
In the current version of the Upright model, each product gets a relative score from -5 to 5 in each impact category. The best products in an impact category get a 5.0 and the worst get a -5.0.
Example: product label “cigarettes" currently gets a 5.0 in negative diseases impact, as our algorithm suggests science says they cause the most diseases relative to volume of operations.
The link between relative and absolute scores is the volume of operations. We currently use revenue as proxy for volume.
Currently we build the backbone for our model using a database of 80 million scientific papers. This means that we base the answer to “which products should get 5.0 and which -5.0 and how does the distribution look like” on causalities documented in scientific research. We have built a neural network that understands and classifies causalities in natural language. In order to do that, we have first manually labeled more than 33 000 scientific papers and used that as training data for the neural network.
We operationalize each impact to impact terms. For example, to find relevant research on diseases we should search with all relevant disease names. We do the same for all products and services.
We then form all "product term + impact term" pairs and search for scientific papers that mention both of them. We feed the papers into the neural network to figure out if they actually said something about the causality between the product and the impact - and if so, what they said. We do this as many times as there are pairs, and take into account value chain links and product family links as described in “How is the Upright product taxonomy built". The average amount of scientific articles contributing to one product is currently roughly 130 000.
This is pretty complex. The real judge of whether this works or not are the results: do they make sense or not. For this, we use real-world feedback data to sanity check how it correlates with our model’s behaviour. You are welcome to give it your own judgment at our free public crowdsourcing environment.