Our AI work

 This page, which I will update, aims to give some links to some of our AI work

1. AI and productivity.

2. Tweet thread describing our results. 

3. Soumaya Keynes discussing our results. 

 3. Markus Academy video.

Planning and energy costs again

 An extraordinary blog on the costs of the planning system via conservative assessments of ecological damage. My italics. 

1. "The first example shared with me related to the questionable designation of a vast 25,898 hectare area of sea (from St Austell to Gribbin Head in Cornwall) as a Special Protection Area (SPA), at  Natural England's insistence. The basis of the designation was the presence of three types of seabird – all of which are widely dispersed around the UK in low concentrations. ...It was suggested to Natural England that the boundary of the SPA should begin at least 500m out to sea, to ensure that housing developments coming forward near the coast weren't unnecessarily required to assess risk of harm to the SPA's seabirds. This request was refused on the grounds that one of the three species was found within this inshore zone: the Slavonian Grebe. Natural England's estimate was that 15 such birds (1.4% of the total population in Great Britain) might use the 25,898 hectare SPA area. Yet as a result of this designation, whenever a new housing development is proposed along this stretch of coast, a habitats assessment must be prepared and considered by the local planning authority in order to assess the risk that people inhabiting the houses go out boating, encounter one of these protected birds and disturb them. The birds in question are only present in winter. The chances of a winter boat tripper encountering one of these 15 birds was said by the Ecologist to be "infinitesimal." 

2. "Sabellaria spinulosa is a worm that builds and lives in tubes on the seabed. When thousands of these tubes are formed together, a reef is formed. 

In recent advice to offshore wind developers, including Outer Dowsing offshore wind farm, Natural England has advised that compensation for impacts to Sabellaria reef is required for placing rock on seabed even in areas that Sabellaria reef isn't present (see paragraphs 11 and 12 [here]).

The consequence of this is that multiple wind farms are delayed while ecological compensation for Sabellaria reef is secured...Projects suffering from this issue include the trio of Norfolk offshore wind projects: Norfolk Vanguard East, Norfolk Vanguard West and Norfolk Boreas. These projects consented in 2021/2022 by Vattenfall and since sold to RWE have been delayed by approximately two years due to inability to satisfy seabed compensation requirements."

3. breeding Guillemot " Dr Trinder's evidence at a recent Dogger Bank South windfarm DCO hearing. You can listen to Dr Trinder [here] at minutes 9 to 20; then his comments on the precautionary approach at minutes 41 to 49; and his thoughts on Natural England's questionable advice around minute 56. I summarise the upshot below.

To understand the potential impact of the wind farm on guillemot during the breeding season, Natural England looked at the maximum distances that six guillemots flew from the breeding site to forage for food during the study period. The distances travelled by each bird were, respectively: 7km, 9km, 27km, 44km, 65km and 338km. The British Trust for Ornithology (BTO) who undertook the study acknowledge ..the 338km figure to be an outlier, and excluded it when calculating the maximum mean foraging distance for these birds. Their expert view was that it was not a realistic data point for the purpose of assessment.

Natural England ignored this recommendation and (including this outlier) calculated the maximum foraging distance to be 153km (putting the wind farm within range in terms of its impact on foraging birds)....The Dogger Bank South project will likely be required to provide compensation for 719 breeding pairs of Guillemot to replace those the SNCBs believe will die as a result of the construction and operation of the project. This will likely be delivered in the form of a combination of predator eradication programmes, artificial nesting structures (bird hotels) and bycatch reduction measures. The estimated total cost of compensation quoted by Dogger Bank South is £173m, although to this includes compensation for other features of the habitat too.

4. The proposed scale of the East Anglia One North (EA1N) offshore wind farm was slashed by 40%, in order to minimise its impacts on Red Throated Diver (RTD) in the Outer Thames Special Protected Area (SPA). This was despite the fact that the developer’s ecological modelling suggested that there would be little or no impact on adult survival (at worst an undetectable 0.1% of the population) due to bird displacement (birds scared by construction)."

Are Manufacturing Jobs Still Good Jobs? An Exploration of the Manufacturing Wage Premium

 From Baynard, et al, 2022. 

This paper explores the factors behind differences in wages between manufacturing and other sectors. Using data from the Current Population Survey, we find that the manufacturing wage premium—the additional pay a manufacturing worker earns relative to a comparable nonmanufacturing worker—disappeared in recent years and that the erosion of the premium has primarily affected workers employed in production occupations, who experienced a wage decline of 2.5 percentage points since the 1990s relative to other workers in production occupations....

 

While the demographic composition and other worker observables introduce level differences in manufacturing premia, our analysis suggests that they are not responsible for the declining trends. A decomposition of the premium by union membership status reveals that declines have been substantially larger across union members. .... 

. We find that the decline in union membership explains more than 70 percent of the decline in premia since the 1990s for union members, but the declines in unionization rates have not significantly affected non-union premia, which have instead responded to other factors, such as capital intensity. Our findings suggest that the erosion of “good” manufacturing jobs has contributed to the increase in overall wage inequality and could accelerate the decline of the manufacturing sector

 

Scanner data (and some other changes) are coming to the UK CPI

 I'm writing this a week before scanner data is introduced into the UK CPI, which will be for the February 2026 inflation figures.  Here is some information. 

1. Overview of how we use scanner data in consumer price inflation statistics: January 2026. 

Overall: 

We will initially introduce scanner data for around 50% of the grocery market. We currently collect 25,000 prices per month directly from shops by price collectors. We will now instead use approximately 300 million price points derived from sales of over a billion units of products per month, collected directly from supermarket scanners in-store and online. For the remaining 50% of the groceries market, we will continue to manually collect prices in-store and online.

Thinking

Consumer price inflation statistics are commonly described as measuring the change in price of a "fixed basket" of goods and services. Historically, we have identified a sample of basket "items" that are representative of what consumers buy and measured the change in price of items in this basket over time. The basket is "fixed" in terms of the items it contains, the quantity of each item in the basket and the quality of those items. This ensures that we only capture changes in price.

However, the idea of a fixed basket is illustrative, and is not a well-defined economic concept. It relates to two concepts: a "cost of goods index" (COGI) and a "cost of living index" (COLI). A COGI is often considered to align with the idea of a fixed basket index, while a COLI accounts for the fact that consumers may change what they buy to less inflationary goods and services. But both concepts measure changes in price.

 How

For the local collection, price collectors visit outlets in locations across the country, and maintain a stable sample by collecting prices for the same products every month. Scanner data, however, reflect the real world. Product availability changes every month when new products enter the market or old products are discontinued. 

The GEKS-Törnqvist multilateral approach works by calculating all possible combinations of "chain-linked" index series in a 25-month window of data, and then averaging them. A "chain link" is the mechanism we use for connecting indices with different baskets. In this context, it is used to refresh the sample of products to maximise the product matches available.

 Unlike data collected in the field, we know how much consumers spend on different product varieties. This allows us to reflect the economic importance of different products through "expenditure weights". Because each linked index uses a different link month, they will also have different weights. 

Quality

It is important for any price index that the quality of products in the sample is held constant, so that changes in product quality do not affect the measurement of price change.

In the local collection, this is managed through a "matched sample". This is where price collectors aim to price the same products every month; where this is not possible, they follow clear procedures to maintain the comparability of the sample, as described in our Consumer Price Indices Technical Manual, 2019. We treat scanner data in effectively the same way. Each unlinked component of the GEKS-Törnqvist is based on a matched sample of transactions. Additionally, the GEKS-Törnqvist mitigates for biases associated with the introduction of new goods into the market through its multilateral approach. 

 As with the local collection, changes in the underlying quality of a grocery product (for example, changes in the ingredients used) are not explicitly captured through this approach. 

 Summary of impact

In this article, we have described how we deal with common issues in producing consumer price inflation statistics when using scanner data. We give an indication of the expected impact from introducing scanner data into our consumer price inflation statistics in our Impact analysis on transformation of UK consumer price statistics: January 2026 article. These impacts are because of the following factors, acting in combination:

• the use of consumption segments, which means that prices for a broader range of goods are captured (as described in Section 5: The scanner data basket)

• the inclusion of multi-buy and loyalty card discount types, which have not previously been captured (as described in Section 6: Accounting for discounts)

• the use of the GEKS-Törnqvist to construct indices, rather than unweighted methods (as described in Section 3: How we calculate price indices with scanner data)

• scanner-data prices reflect the average price paid over three weeks of the month, whereas locally collected prices reflect the price at a particular point in the month (for more information, please refer to our Introducing grocery scanner data into consumer prices statistics methodology)

• scanner data include transactions from all of that retailer's stores, whereas locally collected data only include a sample of stores

2. Impact analysis on transformation of UK consumer price statistics: January 2026.

The average indicative change to the annual rate between January 2019 and June 2025 from the introduction of groceries scanner data was negative 0.02 percentage points for CPIH and negative 0.03 percentage points for CPI, already controlling for the changes introduced in February 2025.

The headline annual rates of CPIH, CPI and RPI were impacted to one decimal place in 38, 39 and 47 out of 66 months, respectively.

Indicative impacts of scanner data at headline level are moderated by groceries accounting for only 12% to 15% of the CPIH, CPI and RPI baskets by weight, and by continued existing in‑store and online collection for around half of the groceries market.

As expected, indicative impacts were larger for more granular indices, where groceries scanner data have a larger proportion of the total weight.

So little overall impact.  what about on food and non-alcholic beverages?

As seen with CPIH, when groceries scanner data were included, “food and non-alcoholic beverages” pushed the CPI annual inflation rate higher in the majority of months in 2020 and 2021, and pushed the CPI annual inflation rate lower between 2022 and September 2024.

Other changes

6.

introduction of the improved measurement of the UK House Price Index (HPI) and

 

changes to price collection for one-night hotel overnight stays

 

nd computer games.

The UK HPI data are used in the Consumer Prices Index (CPI) and Consumer Prices Index including owner occupiers’ housing costs (CPIH) for surveyors’ fees only, and within five items in the Retail Prices Index (RPI). The surveyors’ fees item has a weight of less than 0.2% in 2025 in both CPI and CPIH. The five items have a total weight of around 12% in RPI. The change will take on the improved monthly imputation method introduced into the UK HPI in August 2025. Further information about this UK HPI improvement is available in HM Land Registry’s About the UK HPI guidance.

The changes to computer games affect the items for computer games bought online and computer game downloads, which have a combined weight of less than 0.2% in each of CPI, CPIH and RPI. The indices for these items can be volatile because of the changes in the composition of bestseller charts from month to month. To reduce the volatility and aid interpretation of the data, prices will be collected twice per month.

Similarly, prices of overnight hotel accommodation can be volatile depending on short-term demand and availability of rooms to price. The 2026 changes mean that a hotel price collected six weeks in advance will be collected for two separate nights each month and a hotel price collected one day in advance will be removed from the basket. These items have a total weight of less than 0.9% in CPI, CPIH and RPI.

Energy prices and inflation

 Marvellous from the Resolution Foundation publishing just after the March Iran conflict

It feels like only yesterday that we experienced a once in a lifetime energy price shock. At least the recent price rises in oil and gas remain both A) of unknown duration and B) well below the size of the energy shock caused by Russia’s invasion of Ukraine in 2022.

For our first chart, we’ve considered how the most recent uptick in prices might affect the energy price cap, with suppliers already pulling fixed-tariff deals. If the highs we saw on Tuesday continued (a big if) the cap would increase by roughly £500, undoing and-then-some the £117 fall coming in April thanks to Government policy.

 

And although the increases in wholesale prices remain much less than those seen after the Russian invasion of Ukraine, the government spent a huge amount to protect consumers from those prices. So while the recent moves don’t bring us back to the price cap level seen in early 2023, they do push it over £2,000. This, combined with the effect of a jump in oil prices since the start of the conflict in the Middle East, would add around 1.6 percentage points to CPI by the summer – just when we thought we’d be back to target!

 

and what are those oil prices at least?  10 years of Brent crude from Trading Econ, the most recent observation being Friday 6th March 2026.

Market concentration and intangibles

 M. Bajgar, C. Criscuolo, and J. Timmis, “ Intangibles and Industry Concentration: A Cross-Country Analysis,” Oxford Bulletin of Economics and Statistics 88, no. 2 (2026): 258–274, https://doi.org/10.1111/obes.12659.

gives us the following 

"we use cross-country firm-level panel data for 11 European countries, plus Japan and the United States, between 2002 and 2017, to construct new measures of concentration at the country-industry-year level. We then link these to country-industry level measures of intangible investment" 

"Our econometric results confirm descriptive evidence (Figure 1) indicating that changes in industry concentration are strongly related to intangible investment intensity, particularly in innovation, data and software. The estimates are relatively large: a 1-standard deviation increase in intangible investment (as a share of value added) is associated with a 1.5 percentage point increase in concentration over the next 4 years.7 This corresponds to about a third of the observed concentration increase in the average country and industry. The relationship between concentration and intangibles appears to be similar across the United States, Japan and European countries. It is robust to instrumental variable (IV) estimation, with instruments based on intangible investment in other countries and changes in policy, that is, R&D tax incentives." 

(Link to figure: Open in figure viewer)

Minimum wages

 1. For what it's worth, my model of the effects of minimum wages is this.

2.  The labour market consists of a covered and uncovered sector. In the old literature, the uncovered was illegal working, here I think of it as being self-employment.

3. A job consists of wage and non-wage payments eg. a wage per hour and "working conditions".  The latter might be: holidays, paid tea breaks, number of breaks, effort at work, how horrible or nice the boss it etc. etc. 

4. When a minimum wage comes along, adjustment is along many margins.  Wages might rise, but employment contracting status might change and non-wage attributes might change.  Once the min wage gets to a certain level, there is no more adjustment avaiable though non-wage attributes and employment might then fall.

5. Here is an example of adjustment of attributes.  From the March 2016 NBER digest.  

In Delivering Higher Pay? The Impacts of a Task-Level Pay Standard in the Gig Economy (NBER Working Paper 34545), Yuan An, Andrew Garin, and Brian K. Kovak examine Seattle's App-Based Worker Minimum Payment Ordinance, which took effect in January 2024. The ordinance, which applies to tasks starting or ending within the city limits, establishes a minimum base compensation for delivery tasks as the greater of $0.44 per minute plus $0.74 per mile, or $5 per task. 

The researchers analyze unique data from Gridwise, a third-party app that tracks detailed gig work activity across multiple platforms. Their dataset covers over 2.8 million tasks completed by nearly 6,000 workers in Washington state from August 2023 through July 2024. It includes task-level information on base pay, tips, bonuses, locations, and timing. The researchers compare workers whose pre-reform delivery activity was concentrated in Seattle with workers active elsewhere in Washington state.

The minimum pay standard immediately doubled average base pay per task in Seattle from $5.37 to $12.52 while pay rates remained constant in the rest of the state. However, average tips per delivery declined substantially following the policy, offsetting over one-third of the base pay increase.

For highly attached incumbent drivers, who completed an above-median number of tasks pre-reform (approximately 20 per month), increased base earnings per task were offset both by decreased tips and by a reduction in the number of tasks completed per month. By February 2024, the number of monthly tasks fell by at least 20 percent. Taken together, drivers’ monthly earnings remained virtually unchanged after the reform. 

6. So the non-wage attributes, here tips, entirely offset the wage rise.  There were other effects for non-incumbents, but overall nothing got better.

7. Of course, we want workers to be compensated well.  But this question for min wage setters, I think, is a different one: do they have the information to set better the balance between wage and non-wage attributes than that prevailing in the market?

 

 

Error and fraud in R&D tax credits

 To look at this, this HMRC report, from 2023, looked at claims in 2020-21.  They say that there has been reforms since then.  Here are the findings from " based on 97% of cases in the MREP with a finalised compliance audit" : 

As the final row shows, 16.7% of claims, representing over £1b are error and fraud.  

An updated table, published in October 2024,  gives more current data, based on estimates

And some examples of this in action are by Dan Neidle, here for example on claiming R&D tax credits from footballer wages.

Returning to the 2023, report, some comments from the report

Analysis of the MREP shows that around half of all claims, by volume, contained at least some element of non-compliance. HMRC found fraud indicators in fewer than 10% of claims examined in the random enquiry programme and these claims accounted for less than 5% of the total value claimed. To be classified as fraud, a caseworker needs to have found evidence that the claimant deliberately set out to misrepresent their circumstances to get money to which they were not entitled.

This indicates that the majority of non-compliance is down to other behaviours. As with other regimes, the term ‘non-compliance’ or ‘error and fraud’ encapsulates this full range of behaviours, from mistakes and failure to take reasonable care through to deliberate non-compliance.

In claims where expenditure was over £1 million, around 75% of claims were fully compliant. In smaller claims the percentage of claims being fully compliant was lower, at between 35% and 64%.

As the size of expenditure decreases, the value of non-compliance expressed as a percentage of the value of the claim increases. In the smallest claims where expenditure was less than £10,000, over 75% of the value of the claim was non-compliant. 

The end of rent sharing in the UK

 Bell, Brian, Bukowski, Pawel & Machin, Stephen (2024) The decline in rent sharing. Labor Economics, 42(3), 683 - 716. https://doi.org/10.1086/724570, links here and working paper here study the extent to which wages are correlated with UK company profits/rents.  Using industry and company data and controlling for exogeneity etc. they find a consistent picture, summarised in their Figure 3, working paper version, below

Before 1999, the central estiamte was that wages were marked up by around 25% of firm value added and 7% of  profits.  After 1999 that figure is below 10% and 1% respectively and insignificantly different from zero.  Similar results are obtained for EU industries using industry level data.  

CPI, RPI and CPIH, summary of differences

 

The final table in Consumer Price Indices, Technical Manual, 2019

https://www.ons.gov.uk/economy/inflationandpriceindices/methodologies/consumerpricesindicestechnicalmanual2019