Random research snapshot
Random research snapshotInterference Multiple randomization designs (JRSS-B)
Curious? See more
- The idea
- In a marketplace, buyers and sellers affect each other, so a change given to some of them can spill over to the rest. This design randomizes buyers and sellers separately, which puts every buyer–seller pair in one of four groups, as in the grid. The paper shows how to estimate direct effects and spillovers from these groups, and how much to trust the estimates.
- Why it matters
- Classic experiments split a single population into two groups. They answer few of the questions that matter in a marketplace, where the two sides interact.
- Read
- Paper pagePDFarXiv
Random research snapshotInterference Multiple-unit switchback experiments (preprint)
Curious? See more
- The idea
- Sometimes an experiment cannot show different versions of a service to different users. One alternative is to randomize items instead, switching each item between treatment and control over time. The paper introduces a family of such schedules, balanced across items and over time, and states when its estimates hold up even if an effect lingers after a switch.
- Why it matters
- Item-based experiments have a poor track record for accuracy. In realistic simulations and on real e-commerce data, these designs give much more accurate estimates than standard item-based designs, with no added bias.
- Read
- Paper pagePDF
Random research snapshotPhD work More for less (Biometrika)
Curious? See more
- The idea
- Scientists often sequence a few genomes in a pilot study and then plan a larger follow-up. This paper predicts how many new genetic variants the follow-up will find, based on the pilot. In the figure, the solid part is what the pilot saw and the dashed part is the prediction, with its uncertainty.
- Why it matters
- With a fixed budget, scientists must choose between sequencing more genomes and sequencing each one more accurately. The method allows the set-up to change between pilot and follow-up, so it can help split the budget between the two.
- Read
- Paper pagePDFarXiv
Random research snapshotNoisy data Online activity prediction (Annals of Applied Statistics)
Curious? See more
- The idea
- Before an online A/B test, teams need to know how many users it will reach, and how fast. This paper predicts the number of new users and the total number of visits from a short pilot. Each row in the figure is a user: the first mark is their first visit, the later marks are return visits.
- Why it matters
- How many users a test includes, and for how long, strongly affects what it can detect. On public data and on past A/B tests, the forecasts are often more accurate than other methods, especially when only a few pilot days are available.
- Read
- Paper pagePDFarXiv
Random research snapshotAgentic AI Mind the Gap (IAAI 2027)
Curious? See more
- The idea
- Software changes fast, and the documentation that describes it falls behind. This paper describes a deployed system of AI agents that drafts answers to support tickets and spots where the documentation is missing or out of date. After human review, the fixes go into the documentation that users and their own AI tools read.
- Why it matters
- Each support ticket becomes a chance to fix the docs for the next reader. In the paper's blinded comparison, answers given after a fix were usually preferred over the earlier ones.
- Read
- Paper page
Random research snapshotAgentic AI Can AI agents simulate A/B test outcomes? (WAB@COLM, CODE@MIT 2026)
Curious? See more
- The idea
- An A/B test shows two versions of something to real users and compares the results. This paper asks whether AI agents, each given a profile of real behaviour, can stand in for the users and predict the result before a real test. It splits the gap between simulated and real results into two sources of error, so that each can be reduced.
- Why it matters
- Every real test uses traffic, effort and weeks of time. On past marketing tests, a basic simulation picked up some signal about the direction of effects but overstated their size. Calibrating it on earlier data reduced that error.
- Read
- Paper pagePDFarXiv
Next snapshot: Multi-armed bandits with network interference →
Random research snapshotInterference Multi-armed bandits with network interference (NeurIPS 2024)
Curious? See more
- The idea
- A bandit algorithm learns by trying: it keeps choosing among options, such as discounts, and moves towards the ones that pay off. Here the units are linked, so each unit's result also depends on the options given to its neighbours, as in the highlighted links. The paper assumes each unit is affected by only a few neighbours, and builds simple regression-based algorithms for that case.
- Why it matters
- In an online store, a discount on one product changes the sales of competing products. The algorithms provably learn well, both when the network of who affects whom is known and when it is not.
- Read
- Paper pagePDFarXiv
Random research snapshotNoisy data Bayesian early termination (CODE@MIT 2025)
Curious? See more
- The idea
- Online experiments usually run for a planned length of time. This paper tracks the probability that the treatment is better as the data comes in, and stops the test once that probability crosses a threshold. The shaded part of the figure is the time saved.
- Why it matters
- A shorter experiment costs less and gives an answer sooner.
- Read
- Paper pagePDF
Random research snapshotInterference Experimental design in marketplaces (Statistical Science)
Curious? See more
- The idea
- In a marketplace, buyers and sellers trade with each other. If only some buyers get a change, their trades can also affect sellers who did not get it, as the highlighted trades show. The paper discusses how to design experiments in marketplaces with this in mind.
- Why it matters
- When trades cross between the groups, a simple comparison of treated and control groups can be misleading.
- Read
- Paper pagePDF
Random research snapshotInterference Measuring direct and indirect impacts (TSMO@KDD, CODE@MIT 2024)
Curious? See more
- The idea
- In a marketplace with several sides, a change made for one side can reach the others. This paper reports on an experiment that randomizes groups of units on more than one side at once. It measures two things: the direct effect on the units that got the change, and the indirect effect that spills over to the units around them.
- Why it matters
- Measuring only the direct effect can miss part of the impact of a change.
- Read
- Paper pagePDF
Next snapshot: Regression adjustments for double randomization →
Random research snapshotInterference Regression adjustments for double randomization (arXiv 2026)
Curious? See more
- The idea
- Multiple randomization designs randomize buyers and sellers at the same time, to handle spillovers in marketplaces. This paper uses extra information measured on the units to make the effect estimates more precise. In the figure, the interval around the effect is narrower after the adjustment.
- Why it matters
- The best adjustments differ from those used in classic experiments, and they can be estimated from the data. In simulations, they make the estimates considerably more precise than simpler approaches.
- Read
- Paper pagePDFarXiv
Random research snapshotNoisy data Covariate adjustment at scale (KDD 2023 workshop)
Curious? See more
- The idea
- An A/B test usually compares the average results of two groups. That simple comparison can miss small effects, even with millions of users. This paper studies estimators that also use other data collected about the users, and evaluates them at scale.
- Why it matters
- On real experiments, the adjusted estimates can be more precise and can correct imbalances between the groups. More precise estimates can shorten experiments and speed up decisions.
- Read
- Paper pagearXiv
Random research snapshotNoisy data Statistical power calculations revisited (CODE@MIT 2025)
Curious? See more
- The idea
- Before a test, teams compute its power: the chance that it detects an effect of a given size. The true size is unknown. This paper weights the power at each effect size by a prior belief about how likely that size is, which gives one number, the expected power.
- Why it matters
- It brings what a team already believes about likely effects into the planning of a test.
- Read
- Paper pagePDF
Random research snapshotNoisy data Demonstration experiments (arXiv 2026)
Curious? See more
- The idea
- Sometimes the goal of an experiment is to show that at least one of several options works, for some group and on some outcome. This paper gives statistical tests for that goal that stay valid when the experiment shifts units towards the promising options as it runs. It also gives a rule for that shift, as in the figure.
- Why it matters
- Adaptive experiments are common in online services, health care, biotechnology and the social sciences. The paper applies its methods to a simulated study of cash transfer programmes.
- Read
- Paper pagePDFarXiv
Random research snapshotPhD work Scaled process priors (JASA)
Curious? See more
- The idea
- Genomic studies want to predict how many new variants more samples would reveal. Popular Bayesian models make this prediction depend only on the number of samples, which is too rigid. This paper proposes a new model, the scaled process prior, whose prediction also uses how many distinct variants have been seen so far.
- Why it matters
- On simulated data and real cancer genomic data, it estimated the number of unseen variants more accurately than popular alternatives. Its uncertainty bands also covered the truth more often.
- Read
- Paper pagePDFarXiv
Random research snapshotPhD work Power maximization for rare variants (LMRL@NeurIPS 2021)
Curious? See more
- The idea
- Rare genetic variants may explain much of how traits and diseases run in families. Each one appears in few people, which makes them hard to study. This paper models how design choices, such as how many people to sequence and how deeply, affect a study's ability to find links with disease.
- Why it matters
- With a fixed budget, the model helps choose a design that gets the most out of the study. The paper gives a practical workflow, with examples on simulated data.
- Read
- Paper pagearXiv
Random research snapshotPhD work Double trouble (LMRL@NeurIPS 2022)
Curious? See more
- The idea
- Genomic data is often collected across several groups of people, such as patients with different types of cancer. From a pilot study, this paper predicts how many new variants a follow-up would find in two groups: in total, and shared by both.
- Why it matters
- Methods built for a single group can predict poorly when the groups differ. The paper proves that a natural extension of a leading single-group method fails, and gives the first predictor that handles such differences.
- Read
- Paper pagePDFarXiv
Next snapshot: Cross-study replicability in cluster analysis →
Random research snapshotPhD work Cross-study replicability in cluster analysis (Statistical Science)
Curious? See more
- The idea
- Clustering sorts similar samples into groups, for example to find new subtypes of cancer. A cluster is more convincing if other studies find it too. This paper reviews ways to check this, and proposes a framework that compares clusters across two or more studies, for the whole sample and cluster by cluster.
- Why it matters
- New cancer subtypes found by clustering can affect how patients are managed, so it matters whether they hold up in other data. The checks work with any clustering method.
- Read
- Paper pagePDFarXiv
Random research snapshotPhD work Independent finite approximations (Bayesian Analysis)
Curious? See more
- The idea
- Some Bayesian models allow for infinitely many features, which a computer cannot store. Two common fixes keep a finite number of them: cut the list off, or build a finite version from independent pieces. This paper gives a general recipe for the second kind, and bounds how far it is from the infinite model.
- Why it matters
- The independent version is easier to derive and to run in parallel. In experiments on real data, it performed about as well as cutting the list off.
- Read
- Paper pagearXiv
Random research snapshotPhD work Hierarchical CRMs in trait allocation (BNP@NeurIPS 2018)
Curious? See more
- The idea
- In trait allocation models, each sample has a set of traits drawn from a large pool. With several groups of samples, a hierarchical model lets every group draw from one shared pool, so groups can share traits. This paper derives what such hierarchical models imply once data has been observed.
- Why it matters
- A shared pool lets each group learn from the traits seen in the other groups.
- Read
- Paper pagePDF
Random research snapshotPhD work The Chicago Police Department dataset (NeurIPS 2021, Datasets & Benchmarks)
Curious? See more
- The idea
- Public records about the Chicago Police Department were released separately, by different sources. This paper links them into one documented dataset, in which the records about the same person from every source are joined.
- Why it matters
- One linked and documented dataset is easier to study than many separate releases.
- Read
- Paper pagePDF
About me
I work as a Senior Applied Scientist at Amazon. I spend most of my time working on causal inference and experimental design for large-scale online systems. In particular, I am interested in settings where data is noisy and where interference is at play, like in online marketplaces. I am also exploring how agentic AI frameworks can integrate into online experimentation pipelines to accelerate and improve decision-making.
Previously, I obtained my PhD in Electrical Engineering and Computer Science from MIT in 2021, working in CSAIL under the supervision of Tamara Broderick, working on Bayesian nonparametric methods for prediction and experimental design in the context of genomics studies. Before joining MIT, I was an Allievo student at the Collegio Carlo Alberto and at the University of Torino. Here you can find a copy of my CV. You can contact me via email at lo [dot] masoero [at] gmail [dot] com.
News
- Nov 2026 Mind the Gap: A Deployed Multi-Agent System for Documentation Maintenance accepted at IAAI 2027.
- Oct/Nov 2026 Can AI Agents Simulate A/B Test Outcomes? (with Stefan Hut) presented at the COLM 2026 Workshop on Agent Behavior and CODE@MIT 2026.
- Sep 2026 Multiple Randomization Designs, published in JRSS-B in January, chosen as the JRSS-B Editor’s pick of the year and presented at the RSS 2026 Conference in Bournemouth (proof by pic).
- Sep 2026 Online Activity Prediction via Generalized Indian Buffet Process Models accepted at the Annals of Applied Statistics.
Research
Some of my recent research includes:
Agentic AI for experimentation
Improving efficiency with noisy data
- Demonstration Experiments (2026)
- Leveraging covariate information in A/B testing (KDD 2023)
- Statistical power calculations with effect size priors (CODE@MIT 2025)
- Bayesian early termination for experiments (CODE@MIT 2025)
- Online activity prediction via Indian buffet processes (Annals of Applied Statistics, to appear)
Handling interference in two-sided marketplaces
- Experimental design in marketplaces (Statistical Science, 2023)
- Multiple randomization designs (JRSS-B, 2026)
- Multi-armed bandits with network interference (NeurIPS 2024)
- Multiple-unit switchback experiments (2025)
- Measuring spillover effects in marketplaces (KDD 2024)
- Covariate-adjusted estimators for multiple randomization designs (2026)