iiinova — where your digital self keeps evolving

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

Path signatures capture geometric properties of sequential data that play well with nowcasting. Missing data, mixed frequencies, irregular sampling, non-linearities—all strengths of the method. I hope it goes some way to working in "interesting times", when it really matters!

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

And on a messy data problem with brutally mixed frequencies (with data arriving at fairly odd times): nowcasting UK unemployment using experimental series... yes, even during Covid! This is a month after the end of the furlough scheme, out-of-sample.

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

Simulated data are nice, but the true test is real world problems. And again, we see signatures do a great job—on the canonical nowcasting problem of US GDP, where they outperform dynamic factor models...

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

This all sounds perfect for nowcasting... but does it actually work!? Yes! We show in simulations that regression on signatures does a good job even when lots of data are just missing (cf covid, when stats offices stopped publishing some series) & relationships are non-linear

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

What is the funky stuff? Well, signatures: - handle missing observations and mixed frequencies naturally - play nicely with dimensional reduction - are able to infer nonlinear relationships because they are nonlinear with universal representation properties

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

But guess what? The linear Kalman filter we all know and love for nowcasting is a special case of regression on signatures! So we have a method that does funky non-linear/missing data stuff, but also falls back to a Kalman filter (proof in the paper)

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

A signature is a mathematical object; it describes time-series-like data and is closely related to the theory of rough paths. Basically, it captures geometric information, like increments, in a systematic way. Level 1 terms are changes, 2nd are areas under curves, and so on

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

We wanted something that would work *even when the data are a mess and the target variable is swinging all over the place*. Thanks to Sam Cohen & Lingyi Yang @ox.ac.uk, we had an idea that signatures could be the way to tackle this

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

Doing now/forecasts in covid was hard. But Covid is exactly the kind of period when now/forecasts are most valuable. So I felt 'oh it doesn't work well in "interesting times"' was a cop out, even with missing data issues. Surely we can do better? This paper was motivated by that

100

Arthur Turrell @arthurturrell.bsky.social · 05/10/2026

We just dropped a new version of "Nowcasting using regression on signatures", so time for a 🧵 Take messy, mixed-frequency, gappy economic data, treat it as a continuous path, compute its 'signature', do regression. That's it. But you may wonder: why did we use such heavy mathematical machinery?

140

Arthur Turrell @arthurturrell.bsky.social · 17/09/2026

www.ft.com/content/560f... Great to see this come out—I know incredible levels of hard work have gone into it. More broadly, it's... quite the change in narrative of what has happened to the UK economy over the last 2 decades. May suggest different policies too (targeting hours)

ft.com

UK productivity grew twice as fast as thought after 2008 financial crisis

New official figures put Britain more in line with peer countries in terms of output per hour worked

031

Reposted by Arthur Turrell

John Holbein @johnholbein1.bsky.social · 15/09/2026

There’s a lot of doom & gloom around how AI is affecting higher education. To a certain extent, I get it. AI is making some things harder. But I’ve found that AI has in some ways greatly enhanced my teaching. For one, it’s opened up applied activities that weren’t possible before. One example: 🧵

26210

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

Huge team involved in this work, as you can imagine: those I have handles for include @mihaicod.bsky.social , @danielrock.bsky.social, @aleximas.bsky.social, @jmateosgarcia.bsky.social... full list in the paper: ai.google/static/docum...

ai.google

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

So what's our early take on AI as an innovation in the method of innovation? There's significant potential to increase scientific productivity, & 7 hours/week (big!) is *already* being saved. But it's early days & bottlenecks (inc. verification) limit sci-productivity take off

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

looking a bit beyond, we expect time spent on bottleneck tasks to grow as a %, more papers that recombine ideas from across disciplines, & publication volume increases for fields that don't need wet lab work. Unfortunately, not all of this volume may be high quality ("sci-slop")

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

There's another effect too though—as some things (hypothesis generation, modelling) become 'cheap', science gets bottlenecked by other tasks... 7) 89% of those who save time spend more than a tenth of it checking AI outputs & 41% report a growing backlog of untested hypotheses

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

But are we getting *improvements*? 6) 74% of scientists surveyed report net time savings; (6% report losses). Average saving of ~7 hours/week. This is mostly going back into more research. 47% report an increase in outputs of 10% or more

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

4) specialised models and LLMs are complements: LLMs do code, stats, literature, drafting, teaching, admin while specialised models predict, generate, classify — eg protein structures, weather. 5) within LLMs, analysis & modelling + communicating research findings are top 2 tasks

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

1) In our data, scientists are about 2.7x overrepresented in AI use compared to what you'd expect from employment share. Scientists are using AI a lot! 2) Science convos are longer (tokens), more multimodal, and involve more domain expertise 3) Computer science has highest use

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

Where might we see innovation in the method of innovation first? Well, science of course—the engine of innovation and one of the main drivers behind the huge increase in prosperity over the last 200 years. So what do we find?

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

First, why does this matter? If AI is a tech boost we get a one-off economic growth boost. But if AI is an "innovation in the method of innovation", we get long run higher expected future growth rates! Compound interest makes those scenarios very different for future prosperity

110

Arthur Turrell @arthurturrell.bsky.social · 15/09/2026

Today, my Google colleagues and I published work led by @mihaicod.bsky.social on AI in Science, drawing on 15 million Gemini interactions, 2,600 specialised AI models & a 600-scientist survey. What did we find about how AI is changing science? Read on...

131

Arthur Turrell @arthurturrell.bsky.social · 04/09/2026

How can AI help statistics? And why, with everything going on, should you care? "Statistics and Stupidly Smart AI" was filmed @rigb.org & based on my year visiting schools to talk stats for @royalstatsoc.bsky.social You'll be amazed at what AI can do from a laptop! www.youtube.com/watch?v=G0CY...

youtube.com

Can AI fix Britain's broken statistics? | with Arthur Turrell

YouTube video by The Royal Institution

020

Arthur Turrell @arthurturrell.bsky.social · 30/07/2026

Incredible that you picked the two places I have spent the most time in my life. I just thought phone signal was dodgy everywhere!

020

Arthur Turrell @arthurturrell.bsky.social · 23/07/2026

Our team has today released a huge analysis of the AI economy as seen through 15 million aggregated and de-identified human-AI interactions across the Gemini App, AI Mode, and the Gemini API. Most comprehensive look at how real people are using AI at scale yet! blog.google/innovation-a...

blog.google

Understanding the AI economy

We’re releasing the first Activity, Task, Landscape, and Adoption Study (ATLAS) report, showing how people use Google’s AI tools.

341

Arthur Turrell @arthurturrell.bsky.social · 14/07/2026

After over a decade (!) in the public sector, at the Bank of England, No10, the ONS, and even the Fed Board, I'm starting a new role at Google working on the economics of AI. For me, how AI will reshape the economy is the biggest and most exciting topic in the world right now! Watch this space!

020

Arthur Turrell @arthurturrell.bsky.social · 13/07/2026

Press on the no10 Innovation Fellows scheme, which I have just left. To improve the UK, we need to bring a wide range of talent into gov. I was stunned by the commitment, energy, and talent of my colleagues on this scheme: every dept should have one! www.ft.com/content/4df5...

ft.com

The ‘innovators and disrupters’ hired to bring AI to the UK public sector

Government fellowship scheme places tech experts on ‘high-impact tours of duty’ to improve services

021

Arthur Turrell @arthurturrell.bsky.social · 11/07/2026

I forgot to mention... there is air conditioning in the lecture hall AND we'll be finished well before the football 😉

000

Arthur Turrell @arthurturrell.bsky.social · 10/07/2026

Over the last 12 months, I've been visiting schools as a @royalstatsoc.bsky.social William Guy Lecturer talking about statistics & how AI could help improve them Tomorrow, 7pm, I'll be @rigb.org to do the same with demos, stats, & LOTS of AI Get a ticket & come along! www.rigb.org/whats-on/sta...

rigb.org

Statistics and stupidly smart AI

Statistician Arthur Turrell explores how we can separate bad data from good, how it impacts decisions, and w

121

Reposted by Arthur Turrell

Royal Statistical Society @royalstatsoc.bsky.social · 08/07/2026

We're really looking forward to this event with @arthurturrell.bsky.social, our brilliant William Guy Lecturer and member of the RSS Education Advisory group

052

Reposted by Arthur Turrell

Royal Institution @rigb.org · 22/06/2026

How can AI improve statistics and take the pulse of our nation in new ways? Statistician @arthurturrell.bsky.social will illuminate how AI can be used in a considerate manner, for better data.

rigb.org

Statistics and stupidly smart AI

Statistician Arthur Turrell explores how we can separate bad data from good, how it impacts decisions, and w

153

Arthur Turrell @arthurturrell.bsky.social · 03/06/2026

Come and work @downingstreet as a data scientist! We're hiring for a data scientist working on health and education. You'll be right at the heart of UK policy making and (IMO) working with the best tech stack in the UK public sector. Link: www.civilservicejobs.service.gov.uk/csr/jobs.cgi...

civilservicejobs.service.gov.uk

Quick Check Needed

2817

Reposted by Arthur Turrell

Dariia Mykhailyshyna @handle.invalid · 01/05/2026

Reminder that this workshop by @arthurturrell.bsky.social has been rescheduled to next week (May 7) and you can still register! #EconSky #AcademicSky

095

Arthur Turrell @arthurturrell.bsky.social · 15/04/2026

This looks very nice indeed, excited to try.

020

Arthur Turrell @arthurturrell.bsky.social · 10/04/2026

This looks like a spectacular public good #econsky (Full disclosure: @packlesshepherd.bsky.social is a co-author but my main skin in the game is maybe we'll finally finish our paper together now...) github.com/Davidvandijc...

github.com

GitHub - Davidvandijcke/coarse

Contribute to Davidvandijcke/coarse development by creating an account on GitHub.

0114

Arthur Turrell @arthurturrell.bsky.social · 04/04/2026

I feel that the integration with calendars should also be possible in today's world, but maybe there's not a widely used open protocol? It's a fair point on attack surface

100

Arthur Turrell @arthurturrell.bsky.social · 03/04/2026

For me it's that it's unstable and crashes a lot. My experience and that of everyone I know who has used enterprise versions.

100

Arthur Turrell @arthurturrell.bsky.social · 03/04/2026

Obviously a link in to Artemis II and its recent Outlook snafoo!

110

Arthur Turrell @arthurturrell.bsky.social · 03/04/2026

Given building software is now cheap, asked someone what would make biggest difference to their public sector job: "Outlook not crashing all the time". Enterprise Microsoft IT people: do you outright block other front-ends (eg Thunderbird) or would it be possible to have both in secure envs?

430

Reposted by Arthur Turrell

Dariia Mykhailyshyna @handle.invalid · 20/03/2026

❗️Our next workshop will be on March 26th , 6 pm CET on Accelerate your research: reproducible and efficient analysis in Python by @arthurturrell.bsky.social Register or sponsor a student by donating to support Ukraine! Details: bit.ly/3wBeY4S Please share! #AcademicSky #EconSky

194

Arthur Turrell @arthurturrell.bsky.social · 24/03/2026

sites.google.com/view/dariia-... Hey everyone, how'd you like to learn about reproducibility and efficiency with Python code AND support Ukraine? Well, on Thursday evening at 18:00 - 20:00 CET you can! For just 20€ (more if you like) you can join an online class. More details at the link

sites.google.com

Dariia Mykhailyshyna - Workshops for Ukraine

Feedback on the past workshops (if you want to learn how to make wordclouds, check out Text Data Analysis workshop below)

123

Reposted by Arthur Turrell

Alex @s3alfisc.bsky.social · 23/03/2026

I am very excited that PyFixest 0.50.0 is on PyPi, including a new graph-based solver for demeaning that makes fixed effects estimation in PyFixest significantly faster for "sparse" fixed effects structures.

1157

Arthur Turrell @arthurturrell.bsky.social · 19/03/2026

Relentlessly prune the older ones! I tend to only really use one frontier model (RAM heavy) and one much smaller model for simple tasks (RAM light) at any time. Which models those are changes over time as better ones are released but personally rarely need a third.

101