Interpretable Anomaly Detection for Tokenised Real Estate and Real-World Asset Markets in the UAE

A SupTech Toolkit Anchored to AUS, MBZUAI, and the Live Dubai Land Department / VARA / Prypco Pilot.

10-word hook. Interpretable anomaly detection for tokenised UAE real estate, with regulators.

Working title for the EOI submission. The full 22-page proposal in May will refine the title.

Abstract

The Dubai Land Department, Dubai’s Virtual Assets Regulatory Authority, and Prypco Mint launched the UAE’s first live tokenised real-estate pilot in May 2025; the Securities and Commodities Authority’s Decision 15/RM/2025 then created a federal legal framework for security tokens and commodity-token contracts in June 2025. Tokenised real-world assets in the UAE crossed roughly seventeen billion US dollars on-chain by the end of 2025, with real estate leading. As Phase-1 issuers extend the framework toward tokenised commodities through 2026 and 2027, supervisors face an immediate gap: there is no open SupTech tooling tailored to the UAE token taxonomy that detects manipulation, oracle decoupling, or tail-risk co-movement events with explanations a supervisor can use. This proposal builds that toolkit. It develops rwa-suptech, an open-source library that ingests on-chain transfers, oracle reference prices, off-chain news, and exchange order-book deltas; outputs alerts using interpretable extreme-value-theory anomaly detection; attaches counterfactual explanations generated from a diffusion model over the input feature space; and exposes a regulator-readable API. The methodology builds on the team’s joint precedent of FRG23 (“Medium Anomaly and Fraud Detection in Blockchain and Cryptocurrency Networks”, AUS, 2023-2025) and FRG24 (“From Digits to Dollars: Evolution of Price Impact in Digital Assets”, AUS, 2024-2025), which together produced the Physica A 2024 stylised-facts paper and a joint blockchain-anomaly primer with working code. The deliverable arc starts at Applied (month 0), reaches Tech Development by month 18, and demonstrates a first Validation step by month 24 with at least one named regulator-side or exchange-side analyst actively using the toolkit on real or pilot data. Primary domain is tokenised real estate; tokenised commodities (gold, oil) are a Year-2 extension contingent on Phase-1 issuer availability. Capacity building is anchored at the American University of Sharjah, the Mohamed bin Zayed University of Artificial Intelligence, and the COST Action 19130 European regulator advisor network.

Specific aims

  1. Methodological aim. Develop an interpretable extreme-value-theory anomaly detector with counterfactual explanations for tokenised real-world asset markets, building on FRG23 / FRG24 methodological precedent. The novelty is the integration of multivariate tail dependence, online change-point detection, and diffusion-model counterfactuals into a single regulator-readable pipeline.
  2. Data aim. Produce the first open dataset linking tokenised real-estate transactions (Prypco / DLD pilot) to underlying-asset reference signals (rental yields, comparable transaction prices, location indices), released CC-BY at end of year 1.
  3. Software aim. Release rwa-suptech as an open-source MIT-licensed library with alert API and explanation dashboard, validated end-to-end.
  4. Validation aim. Validate rwa-suptech on at least two real or pilot data streams from the UAE RWA market by month 24, with a named user (a supervisor at SCA, ADGM RegLab, DFSA, or a risk analyst at DLD / Prypco). The named-user commitment is a Year-2 milestone; Year-1 validation runs against AUS-internal substrate plus public on-chain data.
  5. Capacity aim. Train one UAE-based PhD recruited through AUS Office of Graduate Studies; launch a graduate-level course “Interpretable AI for Tokenised Markets” at AUS Mathematics & Statistics in semester 2 of year 2; host two workshops, one at AUS in year 1 and one at MBZUAI in year 2, open to UAE researchers.

Methods and approach

The pipeline takes four input streams: on-chain token transfers, oracle reference-price feeds, exchange order-book deltas where available, and off-chain news arrival timestamps. The anomaly-detection layer uses unsupervised methods consistent with Stephen Chan’s distribution-theory and EVT publication record: bivariate Peaks-Over-Threshold for tail dependence, regime-aware extreme-value distributions, and online Bayesian change-point detection to flag regime transitions. The counterfactual-explanation layer uses a diffusion-model approach over the input feature space, returning a regulator-readable text artefact alongside each alert. We rely on Joerg Osterrieder’s 2025 reaction-times-to-macro-news methodology to align off-chain news arrival times with on-chain anomaly events as auxiliary explanatory features.

Two clarifications on scope. First, this is interpretable anomaly classification with counterfactual explanations, not causal-identification econometrics. The team has not published causal-ID work; the proposal does not claim treatment-effect identification. The counterfactual-explanation generator is a descriptive interpretability tool, not a treatment-effect estimator. Second, foundation models are used as off-the-shelf inference services for time-series baselines (e.g. TimesFM, Lag-Llama via HuggingFace) rather than fine-tuned components; the team’s contribution is the EVT and explanation layer, not foundation-model training.

The methodology is grounded in two AUS-funded methodological precedents. FRG23 “Medium Anomaly and Fraud Detection in Blockchain and Cryptocurrency Networks” (PI Stephen Chan, AED 248,000, June 2023 to May 2025) produced peer-reviewed papers and smaller code modules and demonstrated the team’s working collaboration through to publication on the Physica A 2024 paper “Stylized facts of metaverse non-fungible tokens” (Chan, Chandrashekhar, Almazloum, Zhang, Lord, Osterrieder, Chu). FRG24 “From Digits to Dollars: The Evolution of Price Impact in Digital Assets” (PI Stephen Chan, AED 25,000, June 2024 to May 2025) added price-impact methodology to the team’s portfolio. Together, FRG23 and FRG24 are the methodological precedent for the proposed work: papers, smaller code modules, and a demonstrated capacity to deliver. The new project will build the open rwa-suptech codebase from scratch under MIT license, applying methods refined through FRG23 and FRG24, rather than lifting the existing AUS-internal code. This avoids overstating the openness of FRG23 / FRG24 deliverables while still leveraging their methodological output.

Application-readiness arc

The project starts at the Applied stage and ends at Tech Development with a first Validation step.

Window Stage Deliverables
Months 0-6 Applied baseline FRG23 / FRG24 methodological summary; new MIT-licensed rwa-suptech repo skeleton; AUS Office of Graduate Studies PhD recruitment opens (target start month 6 to align with AUS academic calendar); first scoping conversation with Prypco and DLD via AUS administrative referral; first cold-email outreach to SCA Securities Tokens unit
Months 6-12 Applied prototype Interpretable EVT-anomaly detector prototype on synthetic + public on-chain real-estate data; first AUS workshop “Interpretable AI for Tokenised Markets - Foundations” in year 1, semester 1; benchmark dataset version 0 released; PhD recruited and onboarded
Months 12-18 Tech Development First real-data validation against Prypco / DLD pilot data subject to access agreement, or against AUS-internal extension if pilot data is restricted; benchmark dataset version 1 released CC-BY; three working papers submitted (one to Risk.net editorial chain, one to Frontiers in AI in Finance, one to Digital Finance Springer); PhD passes thesis proposal
Months 18-24 Tech Development with first Validation At least one named user actively using rwa-suptech on real or pilot data with measured detection latency and explanation quality scored by the user; MBZUAI co-hosted workshop “SupTech for Tokenised Markets” in year 2; ADIA Lab Symposium 2027 talk; MIT release of full library

Deliverables and their named users

Each deliverable has a primary user committed to actively touching it by the named milestone, plus a secondary user.

Deliverable Licence Primary user Secondary user
rwa-suptech open library MIT Named UAE supervisor or exchange risk analyst (target: SCA Securities Tokens unit; backup: DLD / Prypco data analyst; further backup: COST 19130 EU regulator participant via Joerg’s network) by month 24 AUS Mathematics PhD students for thesis chapters
Open dataset of tokenised real-estate transactions linked to reference signals CC-BY COST Action 19130 researchers for cross-thesis comparison ADIA Lab internal research team for follow-on studies
4-6 papers CC-BY where journals allow Academic community via Joerg’s editorial channels at Frontiers AI in Finance, Frontiers Financial Risk and Blockchain, Digital Finance Springer, Journal of Investment Strategies, Risk.net Supervisor training material for the named user
AUS course “Interpretable AI for Tokenised Markets” open syllabus AUS Mathematics PhD cohort Visiting MBZUAI students under cross-institutional agreement
Two workshops (AUS year 1, MBZUAI year 2) open recordings, CC-BY UAE early-career researchers Regional regulator analysts

The UAE supervisor commitment is the load-bearing claim. The proposal acknowledges that supervisor access is not yet confirmed; the partner-engagement plan below sets out the cold-outreach pathway in May 2026 and the fallback users if outreach does not materialise.

UAE engagement plan

Partner Status (honest) Specific contact / unit Partner contributes Project contributes
AUS internal in discussion (PI is at AUS) Department of Mathematics & Statistics; AUS Office of Sponsored Research; AUS HPC; AUS Office of Graduate Studies host institution, GPU cluster, PhD/RA pipeline, course host UAE-based open methods, cross-departmental teaching, PhD supervision capacity
MBZUAI confirmed via Co-PI Youcheng Sun MBZUAI Department of Computer Science; Youcheng Sun Assistant Professor Co-PI presence, year-2 workshop venue, MBZUAI-based UAE PhD co-supervised by Youcheng verification and AI-safety leadership for the rwa-suptech alert layer; open SupTech reference for the MBZUAI digital-economy agenda
SCA Securities Tokens unit target outreach during full proposal phase, May 9 to May 29 Director of SCA Securities Tokens unit established under Decision 15/RM/2025; cold outreach via AUS administrative referral Phase-1 issuer data feeds (when issuers materialise), supervisory pilot site, validation user open SupTech tooling tailored to SCA framework, open dataset for the unit’s internal training
ADGM FSRA RegLab target outreach during full proposal phase, May 9 to May 29 RegLab Sandbox programme manager regulatory sandbox access, tokenised-RWA test issuers (Plume Network and similar ADGM-licensed issuers), validation user tooling validated in sandbox, extension to ADGM Islamic-finance tokenised RWAs
DFSA target outreach during full proposal phase Innovation Testing Licence team additional sandbox path tooling extension to DIFC venue
DLD / Prypco / VARA pilot consortium target outreach during full proposal phase DLD pilot programme team; Prypco Mint engineering; VARA Real-Estate Tokenisation working group live tokenised-real-estate transaction data subject to access agreement open analytics layer over Prypco issuances, capacity building for VARA team
COST Action 19130 EU regulator participants confirmed (via Joerg’s chair role in COST 19130) named EU regulator-affiliated researchers within the COST 19130 network of 300+ researchers across 51 countries Year-1 advisor user role: methodology critique, alert-API user testing, cross-jurisdictional applicability review open SupTech reference implementation that COST regulator participants can stress-test in their own jurisdictions

If by May 29 we cannot secure a Letter of Collaboration from any of SCA, ADGM, DFSA, or DLD/Prypco/VARA, the proposal will retain the Year-1 COST 19130 advisor user as the primary committed user, with the UAE supervisor user marked aspirational across Year 2. The Translational Impact score will reflect that honestly. We will not fabricate a confirmation we do not have.

Knowledge transfer plan

All code is released under MIT, all methods under CC-BY, all datasets under CC-BY where licence allows; code is released monthly on a rolling basis; methods are released at the time of paper submission; datasets are released at the end of year 1 with a quality-assurance pass.

Three UAE-anchored capacity-building artefacts:

  1. AUS graduate course “Interpretable AI for Tokenised Markets” at AUS Mathematics & Statistics, optionally interdisciplinary with AUS Computer Science, launching in semester 2 of year 2. The course covers the EVT and counterfactual-explanation methods and uses the project’s open dataset and rwa-suptech library as teaching material.
  2. Two workshops, one at AUS in year 1 (Foundations) and one at MBZUAI in year 2 (SupTech for Tokenised Markets), open to UAE researchers and PhD students. The 2026 ADIA Lab Symposium talk is a separate mandatory awardee deliverable and does not count toward this requirement.
  3. Two UAE-based PhDs, one recruited through AUS Office of Graduate Studies under Stephen Chan and one based at MBZUAI under Youcheng Sun with Stephen as co-supervisor, plus 0.5 FTE AUS RA. Measurable training outcomes: at least three thesis chapters per PhD, two lead-author papers per PhD, and two conference talks per PhD across the project window.

Beyond the UAE, dissemination is led by the COST Action 19130 (300+ researchers, 51 countries, chaired by Joerg Osterrieder), the MSCA-DN Industrial Doctoral Network on Digital Finance (20 institutions, 100 researchers, coordinated by Joerg Osterrieder), and the editorial reach of Frontiers AI in Finance, Frontiers Financial Risk and Blockchain, Digital Finance Springer, Risk.net, and Journal of Investment Strategies. MSCA-DN scope separation: MSCA-DN funds doctoral training and produces multi-thesis research output across the European industrial doctoral consortium; this proposal funds research extending the team’s anomaly-detection and price-impact methods into the UAE tokenised-RWA SupTech domain not covered by MSCA-DN deliverables. The boundary is documented in the project’s annual progress report to ADIA Lab.

Management plan

PI Stephen Chan holds final scientific decision authority and is the primary liaison to ADIA Lab. Co-PI Joerg Osterrieder co-signs methodology and budget decisions and leads dissemination. Co-PI Youcheng Sun owns the verification and AI-safety workstream and co-supervises the MBZUAI-based PhD. AUS Office of Sponsored Research handles fiscal monitoring against AUS internal procedures and the ADIA Lab Cooperative Agreement reporting cadence. MBZUAI contributions through Co-PI Youcheng Sun are handled via cross-institutional Memorandum of Understanding plus an in-kind PhD funding line. Communication cadence: bi-weekly tri-PI sync of 45 minutes; monthly all-team review including the two PhDs and RA; quarterly partner update to confirmed UAE partners; annual progress report to ADIA Lab plus the mandatory Symposium talk in 2027 and 2028. Decision conflicts are resolved first at the tri-PI sync; if unresolved, escalation goes to AUS Office of Sponsored Research. The risk register and mitigation status are reviewed at every quarterly partner update.

Risks and mitigations

Risk Likelihood Impact Mitigation
UAE regulator (SCA / ADGM / DFSA) Letter of Collaboration does not materialise by May 29 medium-high high on Translational Impact Multi-partner outreach to SCA, ADGM, DFSA, DLD/Prypco/VARA in May 2026; fall back to COST 19130 EU regulator participant as Year-1 advisor user; downgrade Year-2 user from “supervisor” to “exchange or pilot data analyst” if necessary
Tokenised real-estate pilot data access requires non-disclosure or aggregation medium medium Apply differential privacy to derived dataset; use synthetic stress data and AUS-internal substrate during Year 1; restrict open dataset to derived statistics only
AUS PhD recruitment slip beyond month 6 medium medium Pre-identify candidates from AUS spring 2026 applicant pool before May 29 submission; engage Joerg’s MSCA-DN candidate referral as international fallback (subject to AUS visa timeline)
Multivariate-tail-dependence under regime change is technically harder than 12 months allows medium medium Stage methods: univariate EVT in Year 1, bivariate POT and tail-dependence coefficient in Year 2; collaborate with COST 19130 statisticians for methodological review
Compute pressure for foundation-model inference peaks at month 12 and 18 low medium Treat foundation models as off-the-shelf inference; request AUS HPC allocation in advance via AUS IT; cap cloud burst budget
Reviewer slate inadvertently contains COST 19130 / MSCA-DN community members high (because Joerg’s network is wide) medium Hand-curate reviewer slate using the procedure documented in _data/reviewers.yml; explicitly exclude COST 19130 and MSCA-DN affiliated researchers from the suggested slate

Team fit narrative

Stephen Chan brings the EVT and distribution-theory backbone, the AUS UAE base, the FRG23 and FRG24 methodological precedents, the Dubai 4th World Police Summit 2025 talk on Blockchain Forensics, and the Department of Mathematics & Statistics infrastructure. Joerg Osterrieder brings the AI / ML-in-finance methodological breadth, the chair role of COST Action 19130 (300+ researchers across 51 countries) including EU regulator participants, the coordinator role of the EUR 3.8 million MSCA-DN Industrial Doctoral Network, the AI advisory role to ING Group, and the editorial reach across Frontiers, Digital Finance Springer, Journal of Investment Strategies, and Risk.net. Youcheng Sun brings the verification, AI-safety, and formal-methods backbone, MBZUAI Computer Science institutional affiliation, recent publications including Verifi (IEEE TDSC 2024) on verifiable federated unlearning and Isolation-based Debugging for Neural Networks (ISSTA 2024), prior faculty positions at the University of Manchester and Queen’s University Belfast plus an Oxford postdoc, the Google ASPIRE Research Award (2022), and Associate Editor at ACM TOSEM. The three-PI structure spans methodology (Stephen), AI/ML breadth and dissemination (Joerg), and verification and AI safety (Youcheng). The team’s joint Physica A 2024 paper on metaverse NFT stylised facts and joint blockchain-anomaly primer demonstrate prior Stephen-Joerg collaboration to publication. UAE-based personnel beyond the PI: one AUS PhD student funded for 24 FTE-months under Stephen, plus one MBZUAI PhD funded as MBZUAI in-kind contribution under Youcheng with Stephen as co-supervisor, plus 0.5 FTE AUS Research Assistant for 12 FTE-months. Backup PhD recruitment channels via Joerg’s MSCA-DN doctoral candidate pool and MBZUAI’s existing Computer Science admissions cycle.

The team has no prior publications using formal causal-identification methods (synthetic control, instrumental variables, regression discontinuity) and the proposal does not claim treatment-effect identification. The team has no prior hands-on experience training or fine-tuning foundation models with more than one billion parameters; foundation models are used in this proposal only as off-the-shelf inference services. These limitations are honest and bounded: the proposal’s methodological contribution is the integration of interpretable EVT, counterfactual explanations, and tokenised-RWA SupTech, not causal econometrics or foundation-model training. A causal-ML or causal-econometrics consultant or affiliate (target: Athey orbit; one paragraph of advisory commitment) will be approached during the May 9 to May 29 window to strengthen the interpretability layer’s methodological grounding.

Budget shape (US$600,000 over 24 months)

Category Amount Notes
Personnel ~US$420,000 Stephen Chan PI 0.20 FTE per year plus summer; Joerg Osterrieder Co-PI 0.15-0.20 FTE per year; 1 AUS PhD stipend at AUS rate; 0.5 AUS RA
Compute ~US$60,000 AUS GPU shared cluster covers approximately 80%; cloud burst budget sized for two peaks at month 12 and 18
Data ~US$20,000 Reference-price licences (LBMA, ICE for the Year-2 commodity extension), on-chain analytics tooling (Dune, Glassnode subscriptions)
Travel ~US$40,000 PI / Co-PI to ADIA Lab Symposium in 2027 and 2028; Joerg to Abu Dhabi twice per year; PhD to one international conference per year
Subcontracts ~US$0 MBZUAI in-kind PhD funding via Co-PI Youcheng Sun, governed by cross-institutional MoU
Dissemination and workshops ~US$30,000 Two workshops (catering, keynote travel), open-access publication fees, course materials
Indirect (20% of direct costs) ~US$30,000 AUS confirms 20% F&A is acceptable; not a negotiation item

Year-2 extension scope: tokenised commodities

If by month 12 the SCA Phase-1 issuer cohort has launched in volume across DFM, ADX, or other SCA-authorised venues, the project will extend rwa-suptech from tokenised real estate into tokenised commodities (gold, oil) as a Year-2 deliverable. The methodology transfers directly: the EVT layer accommodates new asset classes through configuration; the oracle reference-price stream is replaced with LBMA gold or ICE oil reference; the counterfactual-explanation layer requires no methodological change. If Phase-1 issuance does not materialise in volume by month 12, the commodity extension is dropped and the project completes on tokenised real estate only, without compromising the Year-1 deliverables.

Methodological boundaries

The team has not previously published using causal-identification methods. The proposal therefore does not claim to identify treatment effects. The interpretability layer is descriptive: counterfactual explanations describe what change in inputs would have changed the model’s anomaly score, not what change in inputs would have changed the underlying market outcome (Pearl-style intervention semantics). The reference for the diffusion-model counterfactual technique is AR-Pro (Sun et al., NeurIPS 2024). The team has not previously trained foundation models at scale; foundation models are used only as off-the-shelf inference services through HuggingFace, namely TimesFM and Lag-Llama for time-series baselines and small-scale UNet or transformer backbones (under 500M parameters) for the diffusion-model counterfactual layer. No foundation-model training is in scope. Multivariate tail dependence under regime change is technically harder than 12 months allows; the proposal stages the methodology, with univariate EVT (Peaks-Over-Threshold with Generalized Pareto tail) in Year 1 and bivariate POT plus tail-dependence coefficient estimation in Year 2. The UAE regulator user commitment is contingent on the cold-outreach pathway in May 2026 and the COST 19130 advisor pathway is the named fallback if outreach does not produce a Letter of Collaboration. None of these are blockers; they are honest acknowledgments that calibrate the panel’s expectations to the team’s actual capacity.

MSCA-DN and COST 19130 scope separation

The Co-PI Joerg Osterrieder coordinates the EUR 3.8 million MSCA-DN Industrial Doctoral Network on Digital Finance (20 institutions, 100 researchers) and chairs COST Action 19130 Fintech and Artificial Intelligence in Finance (300+ researchers, 51 countries). Both networks are explicitly scope-separated from the present proposal:

  1. MSCA-DN funds doctoral training at industry-academic partner institutions across Europe; its deliverables are PhD theses and joint training programmes, not a UAE-anchored open SupTech library. The proposed work extends the team’s anomaly-detection and price-impact methods into the UAE tokenised-RWA SupTech domain, which is not covered by any current MSCA-DN doctoral project.
  2. COST 19130 is a network instrument funding meetings, training schools, and short-term scientific missions, not research outputs; its role here is dissemination and the named Year-1 advisor user pathway, not co-funding of the research itself.
  3. No MSCA-DN budget, no COST 19130 budget, and no other identified parallel grant funds the work in this proposal. The PI confirms in writing in the full proposal that the proposed work is not currently awarded or under review elsewhere.

What ADIA Lab itself receives

The project produces three artefacts of direct value to ADIA Lab. First, an ADIA-Lab-acknowledged open SupTech reference toolkit that the Lab can cite as part of its digital-economy thrust, link to from adialab.ae, and host a dedicated session on at the 2027 and 2028 Symposia. Second, a curated open dataset of tokenised UAE real-estate transactions that ADIA Lab’s internal research team can use for follow-on studies. Third, an annual workshop format (AUS year 1, MBZUAI year 2) that ADIA Lab can re-host as an annual ADIA Lab event after the project completes, anchoring tokenised-RWA SupTech as a recurring ADIA Lab research community theme.

References (12 working sources)

  1. ADIA Lab Digital Economy Program Solicitation 2026, retrieved 2026-05-06, ADIA_Lab_Digital_Economy_RFP_2026.pdf in this repository.
  2. ADIA Lab Symposium 2025 Day 3 program, https://www.adialab.ae/symposium-2025-day3.
  3. ADIA Lab Advisory Board, https://www.adialab.ae/advisory-board.
  4. Alexander Lipton bio at ADIA Lab, https://www.adialab.ae/bios/professor-alex-lipton.
  5. UAE SCA Decision 15/RM/2025 commentary (Bird & Bird, 2025), https://www.twobirds.com/en/insights/2025/united-arab-emirates/uae-securities-,-a-,-commodities-authority-consults-on-new-security-token-regime.
  6. UAE crypto regulation 2025 overview (VARA, SCA, ADGM, CBUAE), https://complyfactor.com/uae-crypto-regulation-2025-complete-guide-to-vara-adgm-sca-cbuae/.
  7. UAE RWA tokenisation overview and DLD / VARA / Prypco pilot, https://www.cobo.com/post/real-world-asset-rwa-tokenization-in-the-uae-the-future-of-finance.
  8. Stephen Chan AUS faculty page, https://www.aus.edu/faculty/stephen-chan.
  9. Joerg Osterrieder personal site, https://www.joergosterrieder.com/.
  10. COST Action 19130 wiki, https://wiki.fin-ai.eu/index.php/Joerg_Osterrieder.
  11. Joint blockchain-anomaly primer (Osterrieder, Chan, Chu, Zhang), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4317520.
  12. Joint platform “Anomaly and Fraud Detection in Blockchain Networks”, https://digital-ai-finance.github.io/Anomaly_and_Fraud_Detection_in_Blockchain_Networks/.