A Decade of OCOD: Analysing ten years of offshore-owned residential property in England and Wales

Abstract

In the UK, and particularly London, property is a magnet for the global elite, drawn by both the glamour and the consistent rate of return of the housing market. Often, companies in tax or secrecy havens are used as the vehicle for these investments, making understanding this property class challenging. In 2015, the UK government began regularly releasing data on these offshore property investments. This paper analyses the first decade of this valuable but difficult-to-use dataset, first parsing and classifying the data into a structured format, then identifying the major residential trends. Over the last 10 years, the absolute value of residential OCOD property has increased from £64 to £80 billion; however, the relative value has decreased from 2.4 to 1.9 times the national average. London is the hub of offshore residential property, accounting for 45% of all properties by volume and 81% by value. However, these overall changes obscure more complex dynamics as luxury properties, owned by companies incorporated in the British Virgin Islands, decreased in value and volume across the time period. In contrast, there was an increase in housing developments being built by companies incorporated in Jersey and Guernsey. As Named Entity Recognition using deep learning is a critical part of the address parsing pipeline, we test different approaches and find that although training a model using weakly supervised learning (F1 = 0.97) marginally outperforms a model trained on a small, high-quality dataset (F1 = 0.95), this does not outweigh the cost of creating the labelling rules. As a result, we believe that investing more time in high-quality manual labelling is a better choice. The outputs of this paper are a pip-installable Python library and a pre-trained NER model that processes the datasets, increasing accessibility to this valuable resource for social and economic study.

Keywords

housing NLP machine learning dataset tax

Introduction

London and New York have long been considered the two prime cities of the world. This primacy has been strongly related to their status as finance hubs and has resulted in property in these two cities taking on a significantly financialised aspect (Fernandez et al., 2016). However, in recent years, the profile of the two cities has begun to diverge, as London has started catering to the wealthy themselves, rather than creating wealth (Taylor and Derudder, 2022). In London, the internationally super-rich (Cousin and Chauvin, 2021) own properties which are ‘leisure-related investments’ (Paris, 2009), providing both a return on investment due to London’s robust housing market, and prestige provided by the city. This has led to London being an extreme outlier in terms of the internationalism of the rich that make up the upper echelon of its society (Atkinson et al., 2024a).

Beyond the global elite, buying property for themselves, developers also see opportunities in London, creating property developments for both the global elite and the tier below (Atkinson and Mingay, 2024). Such developments, often using offshore-owned companies (Bourne et al., 2023), can be sold directly or as part of a financial instrument (Aalbers et al., 2023).

This influx of money has not been without issues as London has become a global centre of fraud, money laundering, and corruption (Hollingsworth and Lansley, 2010; Atkinson, 2021; Treasury, 2020; TIUK, 2022; Bullough, 2022; Kuldova et al., 2024). In recent years, the UK government has responded to this issue with additional policy focus (NCA, 2021, Westminster-council, 2022; Tugendhat et al., 2018), and legal changes forcing the creation of registers of beneficial owners to increase the transparency in the offshore-owned property market (UKGov, 2018b). Additionally, in 2021 the UK government applied a 2% stamp duty surcharge to residential property in England and Northern Ireland for individuals who are non-UK residents (UKGov, 2025b).

The impact of property investments by the internationally ultra-rich and developer companies can amount to ‘elite-capture’, and can not only affect the price of property (Badarinza and Ramadorai, 2025; Sá, 2025) but also have a material effect on the makeup of the city, both at the luxury (Atkinson and Mingay, 2024; Atkinson, 2021; Lauermann et al., 2024) and affordable end of the market. In London, such physical manifestations of these luxury investments have been referred to as empty homes (Bourne, 2019), ultraland (Atkinson and Mingay, 2024), and the basement belt (Burrows et al., 2022). Studies at the global level have found that the elite capture of the city can be in part facilitated by a two-tier system of law (Lauermann and Mallak, 2023), which leads to situations where gentrification is critiqued as a form of domination of ordinary citizens (Jenkins, 2025). Ultimately, this can result in social movements aimed at regaining control (Lauermann et al., 2024).

In 2015, the UK government released the ‘Overseas companies that own property in England and Wales dataset’ (OCOD) (Land Registry, 2022), due to a freedom of information request by the satirical political magazine ‘Private Eye’. Since then, the OCOD dataset has been released regularly, and 10 years after its initial release, there are now over 100 editions of the dataset, which can be mined for insight.

The OCOD dataset has been used previously in research to provide insight into the UK and specifically the London housing market (McKenzie and Atkinson, 2020; Johannesen et al., 2022; Bourne et al., 2023; Sá, 2025). This work has demonstrated the value of the dataset in terms of the impact of offshore-owned property on the market (Sá, 2025), government policy impacts (Johannesen et al., 2022), and its relationship with other forms of unconventional property, such as short-term lets (Bourne et al., 2023). However, the challenges of processing the dataset, along with the complexity of handling supporting datasets, have limited the number of in-depth studies that use it. Whilst Bourne et al. (2023) attempted to create a reproducible process using a weakly supervised deep learning approach, the codebase was not robust or flexible enough to be usable to others and failed to increase uptake of the OCOD dataset. In addition, the use of deep learning for OCOD address parsing has been criticised as being opaque and at risk of inducing error (Johannesen et al., 2022), whilst the broader idea of weakly supervised learning has also been brought into question (Zhu et al., 2023). These issues leave a gap for a method that can parse the OCOD dataset, demonstrating both robustness to error and flexibility to fulfil the needs of diverse research projects.

Research questions

This paper reviews all OCOD data for the first decade of its release between 2015 and 2025, and asks the following questions:

• What trends are there in offshore ownership of property in England and Wales over the last 10 years?

• Which areas have seen increases in the number of offshore-owned residential properties, and how does this relate to the relative value of the property compared to the average?

• Which approach provides better parsing performance? A model trained on a large weakly supervised dataset, a model trained on small, high-quality labelled data, or a simple regex rules.

Method

The methods in this paper represent a significant revision to the work of Bourne et al. (2023), increasing the transparency and reproducibility of the code, and extending the analysis from a simple snapshot to a time series of the first ten years of OCOD releases. This section is divided into six subsections: Data, Regex for Addresses, Model Training, Evaluation, Post-Processing, and Analysis.

For more information on the python library enhance_ocod, please see the Supplemental material section 3. enhance_ocod can be installed using pip or from the github repository where the code is available https://github.com/JonnoB/enhance_ocod. The fine-tuned modernBERT model used for NER is available from Hugging Face at https://huggingface.co/Jonnob/OCOD_NER or downloaded using the library (see Supplemental material for example).

Data

The models in the paper are trained on the same OCOD dataset as well as the development and test sets as used by Bourne et al. (2023), which is the OCOD dataset from February 2022. The test set consists of 1000 hand-labelled addresses, while the development set, used for weak-labelling, comprises 2000 hand-labelled examples, both of which are from the same February 2022 release. For the analysis part of the paper, the OCOD data used in the analysis is taken from all 104 releases between October 2015 and October 2025. The ONSPD dataset of postcodes is the February 2025 version. The Valuation Office Authority (VOA) Ratings list of businesses dataset from 2023 is used; The Land Registry Price Paid Dataset is from November 2025. Together, these publicly available datasets are used to create a standardised dataset of property addresses.

The training and test data can be found at https://osf.io/khavm/wiki?wiki=qevmz. All UK government data can be downloaded from the relevant government websites or using the enhance_ocod library.

Challenges with OCOD addresses

One of the key difficulties with the OCOD dataset is that each address is entered as unstructured free text and can contain multiple addresses. For example, the text below shows a real address from a single entry in the OCOD dataset from February 2022.

Ground to ninth Floor Flats being 101-114, 201-214, 301-314, 401-414, 501-514, 601-613 and 701-704 Alaska Building, 101-114, 201-214,301-314, 401-412, 501-506 and 601-605 Arizona Building, 101-114, 201-214, 301-314, 401-414, 501-514, 601-614, 701-708, 801-804, 901-903 California Building, 101-108, 201-208, 301-307, 401-408, 501-508, 601-608, 701-708, 801-808 and 901-903 Colorado Building, 1-4, 101-109, 201-210, 301-310, 401-410, 501-510 and 601-605 Dakota Building, 1-7, 101-108, 201-208, 301-308, 401-408, 501-506 and 601-604 Idaho Building, 102-112, 201-212, 301-312, 401-412, 501-508 and 601-604 Indiana Building, 1-15, 101-116, 201-216, 301-315, 401-416, 501-510 Montana Building, 101-108, 201-208, 301-308, 401-408, 501-506 and 601-604 Nebraska Building, 1-10, 101-110, 201-210, 301-310 and 402-403 Utah Building, 1-10 and 101-110 Boston Building, 1-6, 101-106, 201-206, 301-306, 401-408 and 501-507 Madison Building, Deals Gateway, London.

The address contains several hundred individual flats/apartments spread across multiple floors of multiple buildings within the same development. Notably, no postcode is given for any of the addresses, and the individual properties are implied by the ‘-’. Addresses such as these are beyond the capability of typical address parsers, which typically expect only a single property (Yassine et al., 2020; Barratine, 2017; Delil et al., 2020). As such, the OCOD dataset requires a carefully designed, custom approach to both the parsing process and geolocation. Ironically, in many ways this address is relatively simple as it does not include number filters such as ‘1-12 (odds only)’ and contains only a single location ‘Deals Gateway’.

The address shown is an example of a multi-property. These properties are typically characterised as being part of a newly developed land parcel and are often associated more strongly with specific countries of incorporation, with Jersey representing approximately 50% of all multi-properties (Bourne et al., 2023). The opposite of a multi-property is a single-property, which, as the name implies, only contains a single property. This type of property is typically very expensive and incorporated in the BVI.

Regex for addresses

Although they ultimately employ different approaches to Named Entity Recognition, both Bourne et al. (2023) and Johannesen et al. (2022) utilise Regular Expressions (regex) to identify spans of letters that form the entities of addresses in the OCOD dataset. Regex uses rules to search for complex text strings and is a fundamental part of computer science and NLP. These rules can be simple, such as searching for the word ‘flat’ to more complex searches; for example, the regex matches character-spans such as ‘23-28’, ‘23 to 28’, and ‘23a to 28c’, to those shown in the address example in the previous section.

Each OCOD dataset is close to 100 thousand lines of data and hundreds of thousands of spans of text, making manual labelling infeasible. However, by building up a set of labelling rules, the entities that make up an address can be identified relatively easily, allowing automatic parsing.

Each entity class has its own set of rules, as shown in Table 1. In total, this paper uses 44 regex rules for entity span detection. The individual regex rules can be found in the labelling.ner_spans sub-module of the enhance_ocod library.

Table 1.

Entity classes and their corresponding number of rules.

Entity class	Number of rules
Building name	7
City	1
Postcode	1
Street name	13
Street number	6
Unit ID	12
Unit type	3
Number filter	1

A key issue with regex rules is that they are not guaranteed to be correct, resulting in erroneous labelling. For this reason, it is important to test the quality of the regex rules either individually or collectively to measure their accuracy and identify areas for improvement. The imperfect nature of this labelling approach is why it is commonly referred to as weak learning.

In addition to errors, the regex rules may conflict with two or more rules claiming a text-span. Address spans are exclusive, meaning that no two spans can overlap; as such, some method is needed to resolve any conflicts. Such a resolution can be performed using a Hidden Markov Model, which evaluates the performance of each rule and chooses the most likely span in case of a conflict. However, these models are non-trivial to create. As the main open-source weak-labelling libraries are now deprecated, this paper resolves overlapping spans by implementing a greedy algorithm which simply keeps the longer of the two spans when there is a conflict. This approach, no doubt, impacts labelling accuracy but is conceptually and practically simple, while reducing dependencies on third-party libraries. Furthermore, as the performance of the labelling can be measured using the development set, the model can be improved if necessary. In the enhance_ocod library, overlapping spans are handled by the remove_overlapping_spans function, the pseudo-code implementation is shown in Algorithm 1.

Once spans are completed a json file is produced, the performance of the regex rules can then be evaluated as described in section, be post-processed into a parsed dataset as defined in section, or used as input data to train a deep learning model, as described in the next section.

Model training

With the weakly supervised labelling process complete, models can be trained and compared to identify the most appropriate method for parsing the entire OCOD history. We will train models using a two-way approach; the models will be trained either on the 2000-row development set or the 94k-row set from the February 2022 release of OCOD. In addition, the data will be pre-processed to ensure consistent spacing and removal of unusual characters, or left as is. The pre-processing is to see if text tokenisation improves, and so results in cleaner parsing. These two tests mean four models will be trained.

In contrast to Bourne et al. (2023), we use a ModernBert Model (Warner et al., 2024). This architecture is an update to the BERT family of models (Devlin et al., 2019; Liu et al., 2019; He et al., 2021), making it faster and improving performance. The high-level training details of the model are shown in Table 2. The code for running the experiments can be found in the run_experiments.py file, while the training script is mbert_train_configurable.py. Both are located in the scripts folder of the code repository. Note that although the model requires a GPU to train, it can be run on CPU only.

Table 2.

Training parameters for the ModernBERT NER model.

Parameter	Value
Model parameters	149M
Learning rate	5e-5
Batch size	16
Maximum tokens	128
Epochs	6
Optimiser	AdamW
GPU	24 GB NVIDIA L4

Evaluation

Once a model is trained, it is important to evaluate its performance and overall ability to parse the address using NER. The ModernBERT models and the overall performance of the Regex rules will be evaluated using the F1 score. This approach is common when measuring the performance of multi-class models. The F1 score is the harmonic mean of the model precision and recall; the definitions of all three metrics are shown in equations (1)–(3).

precision = \frac{True Positives}{True Positives + False Positives}

(1)

recall = \frac{True Positives}{True Positives + False Negatives}

(2)

F_{1} = 2 \cdot \frac{precision \cdot recall}{precision + recall}

(3)

To measure the overall performance of the model across all classes, the micro F1 will be used, which is the mean of the F1 scores for each class.

Post-NER processing

After the most performant model is identified, it will be used to perform NER across all OCOD datasets. With the entities identified, the addresses will be parsed into one address per line. During post-processing, the data is geolocated into LSOA and classified. The post-NER processing has been completely re-designed relative to Bourne et al. (2023). These changes have made the process more robust (see following on entity assignment), almost halved RAM usage to approximately 8 Gb, and made the per OCOD file processing 3 times faster, reducing it from 15 minutes to less than 5.

A critical difference between Bourne et al. (2023) and this work is how entities are assigned to construct a complete address. Previously, a tabular approach was used; however, analysis showed that this method led to many of the multi-addresses being incorrectly parsed, resulting in a substantial under-reporting of the total number of offshore-owned properties. In contrast, this method constructs a parse tree or Directed Acyclic Graph (DAG) by assigning a hierarchical value to each entity type. This approach provides much more predictable and intuitive address parsing. As an example, consider the multi-address ‘Units 1-10, Otter House, Flats 20-30 (evens), Beaver Buildings, River Road, Flats 31-39 (odds), 45, Heron Heights, Stream Street, London, NE43 8QZ’; this creates the parse tree shown in Figure 1. The parse tree is a dependency graph of relations between different entities, where each leaf node creates a new address row in the enhanced OCOD dataset. The parse tree enables the conversion of free-text addresses in OCOD into standardised addresses, with one leaf node per row. As can be seen in Table 3, three of the newly created rows contain multiple unit IDs; these are expanded to be one property per row as shown in Algorithm 2.

Figure 1.

Parse tree showing the dependency graph of the addresses entity components. Each leaf node results in a new address row by tracing back to the root; this allows for a standardised machine-readable address to be constructed from the free-text addresses in OCOD.

Table 3.

The address structure of the example shown in Figure 1 has been correctly parsed into four rows; there is no ’street number’, but the relationship is inherited from the next level of the hierarchy.

Unit type	Unit id	Number filter	Building	Street number	Street name	Post code	City
Units	1-10	–	Otter House	–	River Road	NE43 8QZ	London
Flats	20-30	Even	Beaver Buildings	–	River Road	NE43 8QZ	London
Flats	31-39	Odd	Heron Heights	–	Stream Street	NE43 8QZ	London
–	45	–	Heron Heights	–	Stream Street	NE43 8QZ	London

To minimise the scope for error, the approach employs unit tests that cover a wide range of possible address structures. These tests can be found in the tests module in the enhance_ocod library.

For Geolocating addresses within the government’s LSOA framework, postcode lookups are used. Where postcodes are not available, custom gazetteers were created from the Price Paid Dataset. This allowed matching named buildings within a given LAD and matching street names within a given LAD. Although streets can cross LSOA boundaries, testing the gazetteer showed that 80% of the streets were within a single LSOA; as such, this method appears reasonable.

A significant bottleneck in working with the Land Registry Price Paid Dataset is its size. The dataset is available as a CSV, which makes it take up a lot of space and is very slow to load. If the entire dataset is loaded, it also consumes a significant amount of RAM. A key process change in this paper was the load_and_process_pricepaid_data function from the price_paid_process sub-module. This function can be used to pre-process the Price Paid Dataset, separating it into individual years and saving the results as parquet files. This change reduces the size of the price paid dataset by about 80% (at time of writing, from 5Gb to 1.1Gb), and decreases loading speed for any given year by 1-2 orders of magnitude.

Property types will be identified using regex rules applied to the property address field. There will be 15 default rules used; these are shown in Table 4. The rule set has been changed from Bourne et al. (2023), which was time-consuming to apply and difficult to integrate into the new process structure.

Table 4.

Property classification rules.

Rule name	Description
Land plots at start of address	Identifies properties starting with ‘land’ or ‘plot’
Parking and garage spaces	Identifies parking spaces, garages, and car parks
Airspace properties	Properties related to airspace rights
Penthouse/flat/apartment	Clear residential indicators like penthouse, flat, and apartment
Business properties by address keywords	Business properties identified by commercial keywords in address
Land with words before it	Land or plots that have descriptive words before them
Pubs by building name	Pubs identified by building name patterns
Named building crossing multiple addresses	Named building which crosses multiple street addresses but has no sub-units
Addresses starting with part/parts	Addresses starting with ‘part’ or ‘parts’ are typically businesses
Units without building name	Properties with ‘unit’ or ‘units’ but no building name are businesses
Building match indicator	Properties where a business building was matched
Number match indicator	Properties where the street and number of a business was matched
No street match	If there is no business on the street then it must be residential
oa match	If there is no business in the oa then it must be residential
lsoa match	If there is no business in the lsoa then it must be residential

Having parsed the model into a tidy format (Wickham, 2014) and saved the tabular data as parquet files, the data is now ready to be analysed.

Analysis

Beyond looking at overall trends in the counts of the data, understanding patterns in the property prices of residential properties is also valuable. Previous work (Bourne, 2019; Bourne et al., 2023; Atkinson et al., 2024b) used a Bayesian approach to estimating prices by sampling the price paid dataset. However, it was shown in those works that the distribution about the mean is minimal for large sample sizes. Therefore, instead of using an expensive multiple bootstrapping strategy (Efron and Tibshirani, 1993), this paper will use the weighted mean of the MSOA means of the price paid dataset. As such, we will calculate the mean price of offshore-owned residential property using equation (4).

{\bar{x}}_{w} = \frac{\sum_{i = 1}^{n} w_{i} x_{i}}{\sum_{i = 1}^{n} w_{i}}

(4)

where w is the number of offshore properties in the MSOA, x is the mean price of the MSOA, and n is the total number of MSOAs. This approach gives the same mean value but without the error distribution, which is acceptable for this analysis and much more computationally efficient. In order to make sure that the data includes sufficient sales within each MSOA to make a reliable average, we will use the 3-year rolling mean, for each year of data analysed. In addition, when estimating the price of offshore-owned properties, properties which have not been geolocated will be assigned the Local Authority weighted average instead of the MSOA average. Whilst this will not affect the average property value, it will ensure the absolute value is more accurate than if the value is dropped.

Estimating the value of all property is challenging; in this paper, we will use the number of households per MSOA from the 2021 Census and adjust the total counts using the year household estimate dataset. This approach ensures that the total number of properties per Local authority is correct and restricts the error to misallocation within the distribution of those properties. This will allow an approximate estimate for value, which will be accurate enough at the national level, given the error is minimal during the centre of the time period under analysis. Any MSOA that does not match the price paid dataset estimate will be assigned the mean value for the local authority.

It should be noted that the OCOD dataset itself includes price data; however, it is inconsistent (approximately 55% of rows have prices) and difficult to interpret (it is not always clear what the price represents). As such, previous work has used the Price Paid Data set (Bourne et al., 2023; Johannesen et al., 2022) or housing price index data (Sá, 2025).

Results

This section is broken into three areas of analysis. Firstly, the NER models for identifying the key elements of the addresses are compared. Secondly, the success of classifying property type and geolocation is discussed. Finally, the most highly performing model is chosen and used in the parsing process, and the resultant dataset can be analysed for trends across the decade of data.

Model analysis

The first step in exploring the dataset is to identify the model that most accurately parses the address text. We evaluate the performance of the five competing models on the 1000-observation test set. As shown in Table 5, we find that pre-processing data does not affect model performance; however, the weakly supervised model is the top performer (F1 = 0.97), beating the conventional model (F1 = 0.95), whilst the regex rules perform worst (F1 = 0.92). The F1 of 0.97 on the weak-learning model also just about outperforms the original weak-learning model created by Bourne et al. (2023), which had an F1 of 0.96 on the same test set.

Table 5.

Performance comparison of different model configurations.

Model type	Pre-processed	F1	Precision	Recall
Conventional	True	0.95	0.94	0.96
Conventional	False	0.95	0.95	0.96
Regex	False	0.92	0.92	0.92
Weak-learning	False	0.97	0.98	0.97
Weak-learning	True	0.97	0.98	0.97

The clear performance difference leads us to choose the unprocessed, weakly supervised model as our primary model. Using the same test set as before, we test the NER class performance. As seen in Table 6, the NER class performance is strong, with the F1 score above 0.9 in all classes, except for the ‘Building name’ class, which has an F1 score of 0.88. This indicates that the model is well-trained and can be used to parse the addresses in the OCOD dataset reliably. Table 6 shows that due to the variety of address types present, the counts of each class shown in the ‘support’ column are not even; however, there is a sufficient quantity to provide reliable evaluation. When comparing Tables 1 and 6, it is clear that the number of rules is not an indicator of accuracy.

Table 6.

Performance is high across all classes, indicating the model is reliable.

Class	Precision	Recall	F1	Support
Building name	0.86	0.90	0.88	383
City	1.00	0.97	0.99	947
Postcode	1.00	1.00	1.00	768
Street name	0.99	0.96	0.97	1029
Street number	0.99	0.98	0.98	678
Unit id	0.97	0.95	0.96	370
Unit type	1.00	0.97	0.98	488
Micro avg	0.98	0.97	0.97	4663
Macro avg	0.97	0.96	0.97	4663
Weighted avg	0.98	0.97	0.97	4663

Classification and geolocation

The classification of the property types underperformed the method used in Bourne et al. (2023), but was still acceptable, getting only an F1 of 0.83 relative to 0.94 in Bourne et al. (2023) (see Table 7 for details). This poorer performance was driven primarily by weak recall in the business and residential classes. To understand why performance had reduced, we analysed the performance of each rule across the test set, as shown in Table 8. The table showed that while many of the rules worked very well, the rules that identify residential properties based on the absence of businesses in the VOA dataset did not work very well, with the rule assigning a residential class if there was a very poor performer.

Table 7.

Property classification performance shows some challenges with recall.

Class	Precision	Recall	F1	Support
Airspace	1.00	0.93	0.96	14
Business	0.93	0.63	0.76	287
Carpark	1.00	0.96	0.98	26
Land	1.00	0.99	1.00	178
Residential	0.87	0.76	0.81	482
Unknown	0.01	0.11	0.01	9
Accuracy	0.77	0.77	0.77	0.77
Macro avg	0.80	0.73	0.75	996
Weighted avg	0.91	0.77	0.83	996

Table 8.

There is a substantial difference between the performance of the rules used to classify property type.

Rule	Accuracy	Precision	Recall	F1	Sample count
Penthouse/flat/apartment	1.00	1.00	1.00	1.00	217
Land plots at start of address	1.00	1.00	1.00	1.00	173
Named building crossing multiple addresses	1.00	1.00	1.00	1.00	6
Airspace properties	1.00	1.00	1.00	1.00	13
Building match indicator	1.00	1.00	1.00	1.00	8
Parking and garage spaces	1.00	1.00	1.00	1.00	25
Land with words before it	1.00	1.00	1.00	1.00	4
Number match indicator	0.98	0.97	0.98	0.97	57
Business properties by address keywords	0.98	0.96	0.98	0.97	95
Units without building name	0.86	0.73	0.86	0.79	7
No street match	0.82	0.66	0.82	0.73	157
Pubs by building name	0.67	0.44	0.67	0.53	3
Addresses starting with part/parts	0.56	0.31	0.56	0.40	9
oa match	0.38	0.15	0.38	0.21	39
None	0.08	0.72	0.08	0.14	183

Another area to consider is the geolocation of properties in the Local Authority. Analysis showed that a relatively consistent 10% of properties could not be assigned an LSOA, with 75% of these properties being residential. This makes up approximately 10% of the total residential property portfolio. It should be noted that, as the Local Authority is known, the properties can be assigned the Local Authority average instead of the MSOA average, which makes the impact on the overall price distribution relatively small.

Data analysis

With the data parsed and post-processed such that each row of the dataset truly represents one property, we can begin to analyse the trends across the time series.

After processing the data, there are on average 70% more rows with the mean OCOD dataset having around 95k titles whilst the average processed OCOD dataset has 162k properties (see section 1 of the Supplemental material for more detail). Residential properties were by far the largest property type, making up approximately 65% of all identified property types. The change in the number of residential properties was somewhat noisy, increasing from 102k in October 2015 to a maximum of 112k before suddenly dropping in September 2017, with October 2025 having 106k properties. During the period, the collective value of the offshore-owned residential properties increased 35%, from £64 billion to £80 billion.¹ This represents a relative decrease compared to all properties from 2.4 to 1.9 times the national average (using the weighted MSOA mean). This reduction in relative value means that, despite the total quantity of residential properties changing during the period, the fraction of total property value stays relatively constant between 1.2% and 1.4% of the total value of residential property in England and Wales.

The decrease in relative value is driven almost entirely by London, which contains 47k residential properties or 45% of all offshore-owned property in England and Wales, and 81% by value. Another point of note is the sheer concentration of offshore-owned residential property; 50% of the total value is held in just two of the 318 Local Authorities. The London boroughs of Westminster, and Kensington and Chelsea hold 34% and 16%, respectively, which in 2025 was worth an estimated 39 billion.

Taking the analysis slightly deeper, we find that there are different dynamics at play between the multi-properties and single properties. As shown in Figure 2(a), single properties, typically represented by ultra-luxury properties in central London, have been decreasing in number consistently across the last decade. In contrast, multi-property developments, typically new-build housing, have been increasing slightly. Figure 2(a) shows two distinct events in the multi-properties: A sudden drop between September 2017 and February 2018 and an up-tick in February and March 2022. The 2017-2018 drop was entirely related to companies in Mauritius selling or otherwise de-registering properties. In total, there was a reduction of 8620 properties equivalent to a 91% reduction in property counts, and an 83% reduction in titles held by Mauritian companies, with an estimated value of £2.6 billion. This change was triggered by an amendment to the tax treaty between Mauritius and the UK (UKGov, 1981) which meant that property in the UK owned by Mauritian companies would be charged 15% withholding tax (UKGov, 2018a) on property-related dividends. This amendment closed a loophole that had previously allowed property income to be extracted tax-free. The up-tick in February and March 2022 was focused on Guernsey which gained 3555 properties, approximately 10% of its January 2022 total. This occurred at the same time that the ‘Economic Crime (Transparency and Enforcement) Act 2022’ (UKGov, 2022) was passed. However, unlike the Double Taxation Agreement, whose arrival was foreseen, this bill was rushed through the UK parliament in a matter of weeks as a response to the Russian invasion of Ukraine. Due to the short timelines and the lack of date of registration data in the OCOD dataset, it is not clear whether there is a relationship between the two events.

Figure 2.

The left panel shows the residential property counts with consistently decreasing single properties. The right panel shows that the non-residential property is relatively constant across the time period.

Unlike the changes seen in residential properties, the non-residential volumes have been broadly flat, as shown in Figure 2(b). Another point to note is the difference in volume between the offshore classes, with residential property much higher than the other classes. The ‘unknown’ class is approximately 10% of the total dataset, and manual inspection shows it is primarily a mixture of business (44%) and residential (37%) (for more detail, please see section 2 of the Supplemental material).

When analysing the timeseries of relative value shown in Figure 3, we see that the relative value has declined across both single and multi-properties, but with significantly larger decreases for the higher-value single properties.

Figure 3.

There is a substantial difference in the value of single and multi-residential properties in the OCOD dataset. Additionally, it is clear that the relative value of both property types has been decreasing steadily over the decade.

Because of the value, the over-representation of London in the dataset means that the relative value of nested and single properties for the capital is very similar to Figure 2(a); however, exploring the other regions shows that the pattern of Single properties being considerably more expensive than nested properties is not the norm. Figure 4 shows that five of the remaining nine regions have very little difference between the nested and single properties; Yorkshire and the Humber and East Anglia show the opposite pattern to London, with multi-properties being much more expensive in relative terms than single properties. Only the South East, which in many ways is the hinterland of London, exhibits a similar pattern to the capital, with single properties being significantly more expensive than multi-properties. In addition, the properties outside the capital are substantially less expensive, with virtually none of the time-slices exceeding a 20% premium; indeed, several of the regions show values consistently below the local average. More generally, the relative value of both nested and single properties either stays constant or decreases; only the North Region experiences any increase in value, driven by properties in the Newcastle metropolitan area.

Figure 4.

Looking at the non-London regions, in only 2 of the 9 cases are single properties substantially more valuable than multi-properties.

The Crown Dependencies and the BVI dominate as the country of incorporation, making up an average of 67% of all properties (residential and other) across the period, which in October 2025 was worth approximately £54 billion. However, as shown in Figure 5, we find that there have been substantial and sustained changes in the relative popularity of these territories, with Jersey and Guernsey increasing in volume whilst the BVI and Isle of Man have decreased. These changes are related to how the territories are used, with Jersey and Guernsey being primarily used for the incorporation of large-scale developments and multi-property entities, while the BVI and Isle of Man are the favoured territories for incorporating single properties. However, as shown in Figure 6, the average value of the properties in the BVI is significantly higher than the other three territories, although the average value across all four territories has been consistently decreasing across the last decade, reflecting the trends shown in Figure 4. A valuable question is whether there is any flow of properties from the BVI to Jersey, either through existing or new properties, or whether the BVI-based properties are being moved to alternative jurisdictions. However, such an analysis is beyond the scope of this paper.

Figure 5.

The top four countries of incorporation are the Crown Dependencies and the British Virgin Islands, an overseas territory of the UK known for being a Tax and Secrecy Haven.

Figure 6.

The top four countries of incorporation all show consistent decreases in average property value relative to regular properties.

The change in the spatial distribution of offshore-owned properties in the UK may reflect the broader trends in development and land prices. As seen in Figure 7, whilst London is unsurprisingly a major source of growth in offshore properties, this is not spread evenly across the capital, with some areas seeing a substantial reduction in offshore-owned property. Elsewhere in the country, it is generally the urban centres that see the big changes with substantial increases in the cities of Liverpool, Leeds, and Manchester, whilst the West Midlands, around England’s second city, Birmingham, broadly sees reductions, particularly in multi-properties.

Figure 7.

The change hotspots for multi and single properties are usually linked to the major cities of England, where house prices are highest.

Discussion

The ModernBERT models all outperformed the regex rules approach to entity recognition. Interestingly, although the regex approach performed very poorly on the ‘unit id’ and ‘unit type’ classes, with F1 scores of 0.6 and 0.47, respectively, the weakly supervised learning model achieved extremely high scores on these two classes. This performance disparity indicates that the weakly supervised deep learning model captured the underlying data patterns from the noisy observed data, whereas the simple rule-based regex approach could not.

The high performance of the ModernBERT model suggests that, although deep learning may be less interpretable, its results are consistent and of high quality, making it an effective approach for parsing the complex addresses found in the OCOD dataset. The major advantage of regex rules is that they are very fast to apply; however, since parsing each OCOD set with ModernBERT takes only about 3 minutes, regex may be valuable only in certain niche cases.

In relation to the question of whether weakly supervised learning is preferable to using a small, high-quality dataset, the results of the model comparison showed that the weakly supervised model outperformed the development set model, in line with previous research (Ratner et al., 2017; Lison et al., 2020; Fries et al., 2021). However, the advantage provided by the weak-learning approach was marginal and required substantial time to develop and test the large number of heuristics, supporting the findings of Zhu et al. (2023). This suggests that spending more time building a high-quality hand-labelled dataset would yield a better-performing model than a weakly supervised equivalent with the same development time. The only time weakly supervised learning would be valuable for this kind of data is if the OCOD data licence became so strict that training data could not be used. In that case, the weak learning rules could be validated on a non-publicly available dataset and then used to create the training set used to train the NER model. However, with the current UK data policy favouring open data, this is unlikely to be the case.

Regarding the analysis of the OCOD data itself, across both residential and non-residential properties, the total counts are relatively stable over the period. However, this apparent stability masked dynamics between multi-properties and single properties. Multi-properties appear to follow a slightly increasing trend, while single properties are consistently decreasing in volume. However, both the single- and multi-offshore-owned residential properties showed consistent decreases in value relative to the rest of the market over the 10-year period. These changes mean that, although offshore-owned residential property is still an impressive asset class with an estimated value of £80 billion in 2025, the average property value is becoming more representative of the overall market. The overall decrease was driven by reductions in single properties in London, particularly those based in the BVI, a secrecy haven. These reductions have been sustained since 2018, when a law was announced to enforce a beneficial register of owners in British overseas territories (the law came into force in 2020) (UKGov, 2018b). Such a reduction could be seen as the result of the change in the law successfully increasing the cost of illicit activity, supporting the findings of Collin et al. (2025). Two other notable events are visible in the data which we term the Mauritius Cliff and the Guernsey Bump. The Guernsey bump whilst notable (10% increase in properties), has an ambiguous cause. In contrast, the Mauritius cliff (91% reduction) is clearly a result of the amended Double Tax Agreement, showing the dramatic impact that such policies can have.

The results in this paper differ from those of Johannesen et al. (2022), who find that the proportion of offshore-owned properties increases as a fraction of total property value across the period 2015–2020. However, both this paper and Johannesen et al. (2022) report similar percentage values (approximately 1.3% and 1.3%–1.6%, respectively) for the fraction of the total housing market value comprised of offshore-owned properties. The reason for this discrepancy is likely related to the different methodologies used and the biases that they induce. Our method assumes that offshore-owned properties are drawn from the same distribution as other properties in the MSOA; thus, the mean value can be found from the weighted average. However, this sampling approach is likely to underestimate the value of single offshore properties in London, which appear to be part of a sub-distribution with a substantially higher mean value. In contrast to our method, the approach used by Johannesen et al. (2022) uses direct address matching with price databases, ensuring that individual properties have the correct sales price. However, such an approach results in substantially fewer price data points and could induce bias by overestimating the value of the dataset, as matches in the single property ultra-luxury end of the market may dominate at the expense of the substantially cheaper multi-properties. For example, the hundreds of multi-properties shown as an example in section would not be included as they lack a postcode. As a result, our analysis and that of Johannesen et al. (2022) are biased in opposite directions. Given observed trends of more multi-properties and fewer single properties over time, may explain the differing results.

Unfortunately, a deeper exploration of the differences in price and location distributions with Johannesen et al. (2022) is not possible as their code is not available. Future work could compare the two approaches or develop a hybrid approach to leverage the advantages of both methods while mitigating their downsides. The resolution of the disparity between the findings of our own work and those of Johannesen et al. (2022) is important, as being confident in fundamental trends, such as whether the offshore residential property’s fraction of total market share is increasing, is critical for the analysis and evaluation of government policy.

Beyond the differences between our work and that of Johannesen et al. (2022), the key limitations of this paper are related to the geolocation and classification. In certain cases, it can be extremely challenging to classify addresses using only the textual data available in publicly available datasets. Our classification approach left a substantial number of properties unclassified and yielded mediocre precision and recall scores, suggesting an under-reporting of the total volume and value of offshore-owned property, which makes our number a lower-bound estimate. The geolocation also failed in about 10% of cases; however, the impact of this can be mitigated as the Local Authority is known, so the average value can be applied to any residential property without more detailed geolocation. A focus for future work could be to improve these two aspects of the method. The issues of underperforming (Bourne et al., 2023) in the property classification, as shown in Table 7, appear to be a side effect of the new modular processing pipeline, which differs from the Bourne et al. (2023) pipeline in a way we were unable to identify. This issue increases the amount of ‘unknown’ class properties; further work could focus on finding a mechanism to boost the performance to similar levels.

It should be noted that the uncertainty in residential property location and quantity could be resolved if the Land Registry and VOA used the government’s Unique Property Reference Number (UPRN) when providing data. While this is a mandated requirement, the current omission of this data point creates significant and unnecessary limitations for research into property ownership and market dynamics. The consistent use of UPRN would facilitate more accurate and robust analyses, which are critical for effective policy evaluation. For example, matching the UPRN of properties in the VOA’s business lists to the UPRN found in the Land Registry’s Price Paid dataset would automatically identify businesses relative to residential properties, resolving the classification uncertainty.

Ultimately, the most valuable contribution of this paper is the Python library enhance_ocod, which, by handling data collection, cleaning, and processing, substantially reduces the complexity of working with these large, complex datasets. By creating a Python library and reproducible scripts, we ensure that other scholars can not only use our findings but also build upon them using the modular and extensible library. For example, it is straightforward to train a new model, create new regular expression rules, or change the post-processing pipeline. In relation to the substantial difference in the quantity of multi-properties between Bourne et al. (2023) and this work, the role of an open-source library adds confidence, as the tests can be checked and built upon in a transparent manner. Finally, the creation of a library and the resultant flexible pipelines avoid licensing issues related to OCOD and other government data, which may not be covered by the Open Government Licence (UKGov, 2025a).

Conclusion

This paper significantly develops the work of Bourne et al. (2023) both in terms of analysis and methodology.

For the analysis element, the paper extends the exploration of OCOD from a single snapshot to the entire first decade of data releases. In doing so, we find that there are consistent decreases in the volumes of high-value single residential properties, whilst the multi-property residences, which are typically housing developments, have seen a noisily increasing trend. However, both single and multi-property types have decreased in relative value across the decade, meaning they are now 1.9 times the national average, down from 2.4 times the national average in 2015. As a result of these dynamics, the total fraction of property value comprised of offshore-owned residential properties has remained constant. These findings, particularly the decrease in high-value single properties, suggest that anti-money laundering laws may be having an effect (UKGov, 2018b), but that this is being counteracted by changing trends in house building.

In terms of methodology, we find support for the idea that weak learning is not as effective as spending more time labelling a high-quality training dataset, and that deep learning models outperform deterministic regular expression rules. However, the key output of this paper is enhance_ocod, an open-source, pip-installable Python library. enhance_ocod covers all aspects of data acquisition, deep learning, parsing, and post-processing of OCOD data. We hope that this library enhances the accessibility of the OCOD dataset as a resource to advance our understanding of the economic and social aspects of the housing market in England and Wales. At the same time, the library provides the methodological transparency and reproducibility necessary for researchers to build on each other’s findings.

Supplemental material

Supplemental Material - A decade of OCOD: Analysing ten years offshore-owned residential property in England and Wales

Supplemental Material for A decade of OCOD: Analysing ten years offshore-owned residential property in England and Wales by Jonathan Bourne, Andrea Ingianni, and Rex McKenzie in Environment and Planning B: Urban Analytics and City Science.

Footnotes

ORCID iDs

Jonathan Bourne

Andrea Ingianni

Rex McKenzie

Funding

Publication was paid for by transformative agreement between University College London and Sage.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

No new datasets were generated for the purposes of this article. The dataset used in the analysis can be derived entirely from the \verb|enhance_ocod| library and provided code.

Supplemental material

Supplemental material for this article is available online.

Note

Author biographies

Jonathan Bourne PhD, Honorary Researcher, Centre for Advanced Spatial Analysis, UCL, UK.

Andrea Ingianni PhD, Senior Lecturer, Faculty of Business, Law and Social Science, Kingston University, UK.

Rex McKenzie PhD, Senior Lecturer, Faculty of Business, Law and Social Science, Kingston University, UK.

References

Aalbers

Taylor

Klinge

, et al. (2023) In real estate investment we trust: state de-risking and the ownership of listed US and German residential real estate investment trusts. Economic Geography 99(3): 312–335. https://doi.org/10.1080/00130095.2022.2155134

Atkinson

(2021) Alpha City: How London was Captured by the Super-Rich. Penguin Random House.

Atkinson

Mingay

(2024) Arrive and retreat: London’s ultra-prime apartment blocks as dark urban renaissance. City 28(3-4): 495–511. https://doi.org/10.1080/13604813.2024.2331287

Atkinson

Higgins

Bourne

(2024a) The global urban elite: the sources of wealth and residential networks of the super-rich in 10 cities. Finance and Space 1(1): 299–317. https://doi.org/10.1080/2833115X.2024.2378892

Atkinson

Simas Lima

Bourne

, et al. (2024b) Low-Use Homes in Rural and Coastal Areas of the UK: Geography, Impact and Responses. University of Sheffield. https://www.financialfairness.org.uk/en/media-centre/media-centre-news-article/low-use-homes-in-rural-and-coastal-areas

Badarinza

Ramadorai

(2025) International dimensions of housing markets. https://papers.ssrn.com/abstract=5342263

Barratine

(2017) Libpostal: international street address NLP. https://github.com/openvenues/libpostal.Original-date:2015-03-03T00:21:31Z

Bourne

(2019) Empty homes: mapping the extent and value of low-use domestic property in England and Wales. Palgrave Communications 5(1): 1–14. https://doi.org/10.1057/s41599-019-0216-y

Bourne

Ingianni

McKenzie

(2023) What’s in the laundromat? Mapping and characterising offshore-owned residential property in London. Environment and Planning B: Urban Analytics and City Science 50(9): 2430–2451. https://doi.org/10.1177/23998083231155483

10.

Bullough

(2022) Butler to the World. Profile Books.

11.

Burrows

Graham

Wilson

(2022) Bunkering down? The geography of elite residential basement development in London. Urban Geography 43(9): 1372–1393. https://doi.org/10.1080/02723638.2021.1934628

12.

Collin

Hollenbach

Szakonyi

(2025) The end of londongrad? Ownership transparency and offshore investment in real estate. https://pse.hal.science/hal-05046763

13.

Cousin

Chauvin

(2021) Is there a global super-bourgeoisie? Sociology Compass 15(6): e12883. https://doi.org/10.1111/soc4.12883

14.

Delil

Kuyumcu

Aksakallı

, et al. (2020) Parsing address texts with deep learning method. In: 2020 28th signal processing and communications applications conference (SIU), Gaziantep, Turkey, 5 – 7 October 2020, pp. 1–4. https://doi.org/10.1109/SIU49456.2020.9302154

15.

Devlin

Chang

Lee

, et al. (2019) BERT: pre-training of deep bidirectional transformers for language understanding. ArXiv:1810.04805 [cs].

16.

Efron

Tibshirani

(1993) An Introduction to the Bootstrap. Chapman & Hall/CRC.

17.

Fernandez

Hofman

Aalbers

(2016) London and New York as a safe deposit box for the transnational wealth elite. Environment and Planning A: Economy and Space 48(12): 2443–2461. https://doi.org/10.1177/0308518X16659479

18.

Fries

Steinberg

Khattar

, et al. (2021) Ontology-driven weak supervision for clinical entity classification in electronic health records. Nature Communications 12(1): 1–11. https://doi.org/10.1038/s41467-021-22328-4

19.

Liu

Gao

, et al. (2021) DeBERTa: decoding-enhanced BERT with disentangled attention. ArXiv:2006.03654 [cs]. https://arxiv.org/abs/2006.03654

20.

Hollingsworth

Lansley

(2010) Londongrad: From Russia with Cash. 4th Estate.

21.

Jenkins

(2025) Gentrification as domination. Critical Review of International Social and Political Philosophy 28(2): 188–214. https://doi.org/10.1080/13698230.2022.2119523

22.

Johannesen

Miethe

Weishaar

(2022) Homes Incorporated: Offshore Ownership of Real Estate in the U.K. Discussion Paper DP17738. Centre for Economic Policy Research. https://www.cepr.org/publications/dp17738

23.

Kuldova

Østbø

Raymen

(2024) Luxury and Corruption: Challenging the Anti-corruption Consensus. Policy Press. Google-Books-ID: eB7oEAAAQBAJ.

24.

Lauermann

Mallak

(2023) Elite capture and urban geography: analyzing geographies of privilege. Progress in Human Geography 47(5): 645–663. https://doi.org/10.1177/03091325231186810

25.

Land Registry (2022) Overseas companies that own property in England and Wales dataset. https://use-land-property-data.service.gov.uk/datasets/ocod

26.

Lauermann

Alexander

Mallak

(2024) Contesting “VIP urbanism”. Human Geography 17(3): 255–261. https://doi.org/10.1177/19427786241285043

27.

Lison

Hubin

Barnes

, et al. (2020) Named entity recognition without labelled data: a weak supervision approach. arXiv:2004.14723 [cs, stat]. https://arxiv.org/abs/2004.14723

28.

Liu

Ott

Goyal

, et al. (2019) RoBERTa: a robustly optimized BERT pretraining approach. arXiv:1907.11692 [cs]. https://arxiv.org/abs/1907.11692

29.

McKenzie

Atkinson

(2020) Anchoring capital in place: the grounded impact of international wealth chains on housing markets in London. Urban Studies 57(1): 21–38. https://doi.org/10.1177/0042098019839875

30.

NCA (2021) National strategic assesment of serious and organised crime. Technical report, National Crime Agency. https://www.nationalcrimeagency.gov.uk/who-we-are/publications/533-national-strategic-assessment-of-serious-and-organised-crime-2021/file

31.

Paris

(2009) Re-positioning second homes within housing studies: household investment, gentrification, multiple residence, mobility and hyper-consumption. Housing, Theory and Society 26(4): 292–310. https://doi.org/10.1080/14036090802300392

32.

Ratner

De Sa

, et al. (2017) Data programming: creating large training sets, Quickly, arXiv:1605.07723 [cs, stat]. https://arxiv.org/abs/1605.07723

33.

Sá

(2025) The effect of foreign investors on local housing markets: evidence from the UK. Journal of Economic Geography 25(3): 329–349. https://doi.org/10.1093/jeg/lbae043

34.

Taylor

Derudder

(2022) NY-LON 2020: the changing relations between London and New York in corporate globalisation. Transactions of the Institute of British Geographers 47(1): 257–270. https://doi.org/10.1111/tran.12498

35.

TIUK (2022) Stats reveal extent of suspect wealth in UK property and Britain’s role as global money laundering hub. https://www.transparency.org.uk/uk-money-laundering-stats-russia-suspicious-wealth

36.

Treasury

(2020) National Risk Assessment of Money Laundering and Terrorist Financing 2020. UK Government. Technical report. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/945411/NRA_2020_v1.2_FOR_PUBLICATION.pdf

37.

Tugendhat

Austin

Bryant

, et al. (2018) Moscow’s Gold: Russian Corruption in the UK. House of Commons. Technical report.

38.

UKGov (1981) The double taxation relief (taxes on income) (mauritius) order 1981. https://www.legislation.gov.uk/uksi/1981/1121/made.Publisher:Westlaw

39.

UKGov (2018a) The double taxation relief (mauritius) order 2018. https://www.legislation.gov.uk/ukdsi/2018/9780111168660.Publisher:King’sPrinterofActsofParliament

40.

UKGov (2018b) Sanctions and anti-money laundering act 2018. Technical Report. House of Commons. https://www.legislation.gov.uk/ukpga/2018/13/contents.Publisher:StatuteLawDatabase

41.

UKGov (2022) Economic crime (transparency and enforcement) act 2022. https://www.legislation.gov.uk/en/ukpga/2022/10/contents.Publisher:StatuteLawDatabase

42.

UKGov (2025a) Open government licence. https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/

43.

UKGov (2025b) Rates of stamp duty land tax for non-UK residents. https://www.gov.uk/guidance/rates-of-stamp-duty-land-tax-for-non-uk-residents

44.

Warner

Chaffin

Clavié

, et al. (2024) Smarter, better, faster, longer: a modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference. ArXiv:2412.13663 [cs]. https://arxiv.org/abs/2412.13663

45.

Westminster-council (2022) Westminster launches war on dirty money. Westminster City Council. https://www.westminster.gov.uk/news/westminster-launches-war-dirty-money-3

46.

Wickham

(2014) Tidy data. Journal of Statistical Software 59: 1–23. https://doi.org/10.18637/jss.v059.i10

47.

Yassine

Beauchemin

Laviolette

, et al. (2020) Leveraging subword embeddings for multinational address parsing. In: 2020 6th IEEE congress on information science and technology (CiSt), Agadir – Essaouira, Morocco, 5 – 12 June 2021, pp. 353–360. https://doi.org/10.1109/CiSt49399.2021.9357170

48.

Zhu

Shen

Mosbach

, et al. (2023) Weaker than you think: a critical look at weakly supervised learning. ArXiv:2305.17442 [cs]. https://arxiv.org/abs/2305.17442

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

1.96 MB

0.00 MB