Arthur Turrell

Good news here for working with geospatial data:
parquet.apache.org/blog/2026/02...

The trick seems to be to produce software just good enough that someone can justify not getting something better. I see many teams being told they're getting AI and then receiving MS Copilot for coding 😬

Hi Rachel, thanks for the feedback: I'll pass it all on & it's helpful to know that it's coming across like this. FWIW colleagues on the fellowship (excluding myself of course) are extremely talented & delivery-focused, & many have science backgrounds. Perhaps we need to do more to bring that out.

There's a lot of news out there rn, probably so much that you're distracted from thinking about data analysis. As an antidote, why not try out Skimpy, a Python package for data analysis? It has 496 stars on GitHub: we just need a few more to hit 500. That could be you! ⭐️

github.com/aeturrell/sk...

Excited to be a part of the initiative to "Move Fast and Fix Things", announced in Chief Secretary to the Prime Minister's speech today. One measure is an expansion of the No10 Innovation Fellowship, for which we've launched a new website!

fellows.ai.gov.uk

Speech: www.gov.uk/government/s...

Inevitably, it has turned into an interview about Trump instead!

Interested in what we've been up to at Downing Street? Tune in to Today on BBC Radio 4 at 08:10 this morning!

Doyne Farmer & Francois Lafond have done a lovely paper on falling costs of technology and how it matches a pattern; one of the authors of this could be a good shout
lims.ac.uk/documents/pa...

Line chart showing the dramatic decline in the cost of sequencing a human genome from 2001 to 2022, plotted on a logarithmic dollar scale. The vertical axis shows cost per genome in U.S. dollars, ranging from over $100 million down to about $500, and the horizontal axis shows year. Costs start near $100 million in 2001 during the Human Genome Project, fall gradually to around $10 million by 2005, then drop sharply after 2007 with the introduction of next-generation sequencing technologies. By 2011 costs fall below $10,000, reach around $1,000 by 2017, and continue decreasing to a few hundred dollars by the early 2020s. Annotations mark major technological milestones such as Illumina short-read sequencing, Ion Torrent semiconductor sequencing, PacBio real-time sequencing, Oxford Nanopore’s MinION, and completion of an end-to-end human genome map.

The cost of sequencing a human genome has fallen over 100,000 fold in nominal terms since 2001.

In a new visualization, I've added some of the key advances in sequencing during that timeline:

In person!

Lovely finishing plug for @dianecoyle1859.bsky.social's book at #ASSA2026 from the panel on government statistics!

Former BLS commissioner Bill Beach says the agency has lost 25% of its staff since January. 40% of leadership positions are vacant. And DOL leadership “does not seem to support the Bureau.”
#ASSA2026 #EconSky

Hello #ASSA2026 attendees! Interested in better nowcasting? We have a top line-up of speakers on the latest and best ways to gauge the state of the economy in real-time at:

🎤 Next Gen Nowcasting: Signatures, Distributions, and Simplified Workflows

📍309

🕣 Sunday, 14:30 EST

#econsky

Heading to #ASSA2026? Interested in gauging the state of the economy in real-time—even during crises? Then come to...

🎤 Next Gen Nowcasting: Signatures, Distributions, and Simplified Workflows

📍309

🕣 Sunday, 14:30 EST

eppro01.ativ.me/appinfo.php?... #ASSA2026
#econsky

If you're heading to Philadelphia for the ASSA, I'll be speaking about this work: we'd love to get your feedback.

And this approach performs well for nowcasting UK unemployment, **even over rocky periods like the pandemic** when the numbers were subject to all kinds of shocks and uncertainties... eg a month after the end of the furlough scheme

Is it any good at nowcasting though? Yes! On the standard test, nowcasting US GDP, signatures (combined with DFM or PCA) are performant vs sensible benchmark models.

Signatures applied to a non-linear series that is irregularly sampled

We show that signatures generalise linear Kalman filters... so you can retrieve the kind of behaviours/predictions you'd expect. But as well as interruptions, signatures can also incorporate non-linear transformations. They work in "noisy" situations!

Mathematical definition of signature

The key insight in using signatures is that they are well-suited to interrupted paths, and so can comfortably deal with irregular and mixed frequency observations. If time series are temporarily unavailable, or published at a different time or frequency, signatures are cool with this 😎

Chart showing an illustration of two levels of the signature

Now-warned is now-armed, to ruin a phrase: policymakers need to know what's happening. And that's *more* important during stressed times. Lots of nowcasts struggled through the pandemic. Signatures offer a more robust way of nowcasting. Signatures are functions of paths that capture geometric info

Nowcasting using regression on signatures We introduce a new method of nowcasting using regression on path signatures. Path signatures capture the geometric properties of sequential data. Because signatures embed observations in continuous ti...

New working paper: nowcasting using regression on signatures!

We introduce a new method of nowcasting that uses mathematical objects called signatures. Regression on signatures subsumes the linear Kalman filter & is more robust to disruptions in data series 😍

#econsky

arxiv.org/abs/2305.102...

Thanks Alex. Works if it's for data analysis. This is a shorter and more data sciencey version: aeturrell.github.io/python4DS/we...

Russia is ‘exporting chaos’, new head of Britain’s spy agency MI6 warns In her first public remarks since taking office, Blaise Metreweli will also say that UK support for Ukraine is ‘enduring’

"Mastery of technology must infuse everything we do,” she will say. “We must be as comfortable with lines of code as we are with human sources, as fluent in Python as we are in multiple languages.”

New head of MI6

#python

www.ft.com/content/5cbe...

I see, I didn't realise you were actively trying to be rude. I personally really wasn't and I'm sorry for coming across that way.

You just sent me a video of someone mocking something so you must feel it's a valid way of making a point sometimes too? And I appreciate the way what you shared makes the point! NB I understood you were originally talking about any gen AI image, not just slop. Interesting re seniors.

I'm not pretending, I get what you mean! I feel differently & used an exaggerated way to illustrate that but didn't intend to be rude. On broader pt, this is mostly junior researchers & I always think they need some slack. I'm sure my presentations as a junior could have been lot more professional!

They have been in the uncanny valley and I get that it can be distracting if they're obviously wrong. But it's just a tool and I think it will get better, & people will learn to use it better. Personally willing to give some slack to people experimenting with new tools, especially developing ones!

If I can tell someone hasn't hand-painted the graphs on their slides instead of using software, it taints them and makes them look sloppy and cheap.

Those of us giving talks on generative AI 😬

Is Dirk on here?