Hi I’m Bilal, cofounder at https://www.clearbrain.com . ClearBrain is a new analytics platform that helps you rank which product behaviors cause vs correlate to conversion. Think Google PageRank, but for Analytics.
Our founding team worked on this problem for quite a few years while at Google and Optimizely. We contributed to Google Analytics to analyze historical behaviors in seconds, but observing historical trends merely produced noisy correlations. We built Optimizely to measure true cause and effect through A/B testing, but tests took 4-6 weeks on avg to reach significance, and so it would take years to measure the impact of every single page or feature in an app.
So we asked ourselves, could we estimate which in-app behaviors cause conversion, to complement (not replace) a traditional A/B test? We spent a year in R&D, and built ClearBrain as a self-serve “causal analytics” platform. All you have to do is specify a goal - signup, engagement, purchase - and ClearBrain ranks which behaviors are most likely to cause conversion.
Building this required a mix of real-time processing + auto ML + algorithm work. We connect to a company’s app data via Segment, and ingest their app events in real-time via Cloud Dataflow into a BigQuery backend. When a customer uses the ClearBrain UI to select a specific app event as their conversion goal, our backend will automatically run multiple observational studies to analyze how every other app event may cause that goal. This is done in parallel using SparkML, to analyze thousands of different events in minutes. (more on our algorithm here: https://blog.clearbrain.com/posts/introducing-causal-analyti...)
We’ve had beta customers like Chime Bank, InVision, and TravelBank use ClearBrain to estimate which behaviors and landing pages cause their users to convert, and in turn prioritize their actual growth and A/B testing efforts there.
We’re now releasing the product into general availability in partnership with Segment - available on a free self-serve basis today! We look forward to feedback from the HN community. :)
aditiyaa1 1627 days ago [-]
Hi Bilal, Thanks for the overview of the product.
This is a really important business problem to solve for many marketing teams. Just using this to prioritize A/B tests in itself pretty valuable. But one of the concerns around this approach is the un-reliability of causal analysis to estimate true effects. The link below refers to a study done at FB that shows observational studies could be erroneous in estimating effect sizes and in some cases, the direction of the effects. Do you think clearbrain's system is robust enough to estimate the true effects?
Thanks for the great feedback! Yes, some of these limitations expressed in the study are true in the case of ClearBrain - namely we are leveraging observational studies at this time as a prioritized ranking algorithm for which behaviors are most important, but the actual effect sizes themselves may be variable. We're working on improvements, as well as incorporating actual experiment data into our algorithm to make it more accurate over time.
aditiyaa1 1626 days ago [-]
Thanks! What is your strategy around incorporating actual experiment data? Not sure I fully follow here.
newtothebay 1627 days ago [-]
Can you talk about how to infer causality without running an experiment? From your description, "real-time processing + auto ML + algorithm" still sounds very much observational to me.
I'm asking not as knock against your service, but genuine curiosity about how you manage to solve this incredibly hard problem.
EDIT: From your white paper, it looks like you're running a regression that controls for a bunch of confounders. You also interact the treatment variable with those confounders to get the heterogeneous treatment effect.
My concern with that is that we're not controlling for unobservable confounders, which make causal inference so difficult. If we assume that controlling for observable confounders is enough (we shouldn't!), then correlation and causation are the same.
Yep, you're correct that we're using observational studies via a regression to remove confounders and estimate treatment effects. Our confounders are synthetically generated based on the observable variables - we can only make projections of course on digital signals our customers send us (we only use first party data). We are working to incorporate actual experiment data into the algorithm over time as well, to get even closer to the true causal treatment effect.
noobhacker 1626 days ago [-]
Awesome! Could you speak a bit about what you have in mind to incorporate actual experiment data into the algorithm?
pfbtgom 1626 days ago [-]
I think you have an interesting product, but I'm having serious issues with your marketing.
Extraordinary claims require extraordinary evidence. How many of your estimated treatment effects have been supported by experiments? Do you have experiments demonstrating that your model generalizes? How accurate are your estimates compared to experimental results?
It's ironic that you're marketing a causal + analytics product without any data. Generating a narrative and basing it off of observational data is the typical trap that many causal claims fall into. Portraying yourselves as statistical experts and pushing unsubstantiated claims is misleading bordering on unethical.
seanwilson 1627 days ago [-]
In reference to how regular AB testing needs a certain amount of data to get statistically significant results, what kind of level of traffic + conversions would you need for this to work? Would it be useful for sites with low traffic?
bmahmood 1627 days ago [-]
Hi Sean - great point! When I was at Optimizely working on their data science team, we found that on average a test needed 10K-20K unique visitors to reach significance.
We find that this rule of thumb extends similarly to our causal analytics platform. However, we have found that even low-traffic sites are able to get a boost if they are tracking more events on their website (increases the opportunities for signal). Also, our simulations run in minutes on all your historical, rather waiting for weeks for users to be exposed to the test, which speeds up time to insight. If we can not determine significance in our simulation though (due to either sample size or signal), we will designate the projection as a correlation.
gingerlime 1627 days ago [-]
Hi Bilal!
I think I reached out to you in early 2018. Any news about the integration without segment (e.g. Amplitude in our case)? And what’s the pricing model? Couldn’t find much on the site (maybe it’s more limited on mobile?)
bmahmood 1627 days ago [-]
Thanks for reaching out again! We're prioritizing support for Segment at this time, but hope to add other integrations next year. Our analytics product is completely free, so getting set up on our joint solution with Segment shouldn't be too expensive. :)
gingerlime 1626 days ago [-]
Thank you. What’s your monetization strategy then? (Were you acquired by Segment or something? honest question)
polskibus 1627 days ago [-]
How is this related to process mining theory and tools like Celonis?
puranjay 1626 days ago [-]
Nothing about the product but man, that's a fantastically beautiful and informational website. Reminds me of Stripe
bmahmood 1626 days ago [-]
Aww thank you! Our design team is pretty cool. :)
dtran 1627 days ago [-]
Congrats on the launch Bilal and team!
Rendered at 10:44:58 GMT+0000 (Coordinated Universal Time) with Vercel.
Our founding team worked on this problem for quite a few years while at Google and Optimizely. We contributed to Google Analytics to analyze historical behaviors in seconds, but observing historical trends merely produced noisy correlations. We built Optimizely to measure true cause and effect through A/B testing, but tests took 4-6 weeks on avg to reach significance, and so it would take years to measure the impact of every single page or feature in an app.
So we asked ourselves, could we estimate which in-app behaviors cause conversion, to complement (not replace) a traditional A/B test? We spent a year in R&D, and built ClearBrain as a self-serve “causal analytics” platform. All you have to do is specify a goal - signup, engagement, purchase - and ClearBrain ranks which behaviors are most likely to cause conversion.
Building this required a mix of real-time processing + auto ML + algorithm work. We connect to a company’s app data via Segment, and ingest their app events in real-time via Cloud Dataflow into a BigQuery backend. When a customer uses the ClearBrain UI to select a specific app event as their conversion goal, our backend will automatically run multiple observational studies to analyze how every other app event may cause that goal. This is done in parallel using SparkML, to analyze thousands of different events in minutes. (more on our algorithm here: https://blog.clearbrain.com/posts/introducing-causal-analyti...)
We’ve had beta customers like Chime Bank, InVision, and TravelBank use ClearBrain to estimate which behaviors and landing pages cause their users to convert, and in turn prioritize their actual growth and A/B testing efforts there.
We’re now releasing the product into general availability in partnership with Segment - available on a free self-serve basis today! We look forward to feedback from the HN community. :)
https://www.kellogg.northwestern.edu/faculty/gordon_b/files/...
I'm asking not as knock against your service, but genuine curiosity about how you manage to solve this incredibly hard problem.
EDIT: From your white paper, it looks like you're running a regression that controls for a bunch of confounders. You also interact the treatment variable with those confounders to get the heterogeneous treatment effect.
My concern with that is that we're not controlling for unobservable confounders, which make causal inference so difficult. If we assume that controlling for observable confounders is enough (we shouldn't!), then correlation and causation are the same.
White paper: https://blog.clearbrain.com/posts/introducing-causal-analyti...
Extraordinary claims require extraordinary evidence. How many of your estimated treatment effects have been supported by experiments? Do you have experiments demonstrating that your model generalizes? How accurate are your estimates compared to experimental results?
It's ironic that you're marketing a causal + analytics product without any data. Generating a narrative and basing it off of observational data is the typical trap that many causal claims fall into. Portraying yourselves as statistical experts and pushing unsubstantiated claims is misleading bordering on unethical.
We find that this rule of thumb extends similarly to our causal analytics platform. However, we have found that even low-traffic sites are able to get a boost if they are tracking more events on their website (increases the opportunities for signal). Also, our simulations run in minutes on all your historical, rather waiting for weeks for users to be exposed to the test, which speeds up time to insight. If we can not determine significance in our simulation though (due to either sample size or signal), we will designate the projection as a correlation.
I think I reached out to you in early 2018. Any news about the integration without segment (e.g. Amplitude in our case)? And what’s the pricing model? Couldn’t find much on the site (maybe it’s more limited on mobile?)