Every transaction record on Daar originates from the Abu Dhabi Real Estate Centre (ADREC), the government body responsible for registering and publishing every sale, lease, and mortgage in the emirate. We do not supplement ADREC data with listings, broker-reported figures, or third-party estimates. When a number on Daar moves, it moves because the underlying ADREC publication moved — nothing else.

Our data pipeline runs roughly once a day. We compare the latest ADREC figures against our existing core, identify only the genuinely new rows by deduping on a composite key (Date + District + Community + Project + Type + Layout + Price + Area + Rate + Sale Sequence + Asset Class), and append them. We never overwrite previously stored records — preserving our own history lets us compare today’s published figures against what ADREC said about the same transactions last week, catching any restatements or data-quality drift on ADREC’s side.

Post-import, three cleaning passes run automatically. First, a project-name correction list standardises variants like "Sky Gardens Tower", "Sky Tower", and "Sun & Sky, Boutik Mall" into a single canonical name, so volume statistics aren’t split across what is really one development. Second, district names are normalised to a canonical spelling (handling the Al-prefix variance and apostrophe/transliteration differences). Third, transactions below AED 50,000 are flagged as registration fees rather than real sales and excluded from the analytics dataset — this single rule prevents a large number of spurious near-zero "transactions" from polluting district medians.

We do notremove IQR outliers or empty-area rows, which is a departure from what many data pipelines do reflexively. The reason: a high price-per-sqm row might be a legitimate penthouse, and an empty area field might be a pre-handover off-plan transaction not yet measured. Excluding these introduced 5-50% error rates on luxury project deal counts (Four Seasons, Nobu, Faya, and others). IQR filtering is still applied — but at query time, by the specific analytics function that needs it (e.g. rental yield), not at import time. This lets different analyses make different outlier decisions.

Finally, a QA/QC cross-check runs against ADREC’s live statistics API to confirm our aggregate figures match within a 1% tolerance. If they don’t, the import is flagged for manual review before the data reaches the platform.

ADREC’s published residential price index shows roughly +96% from Q1 2020 to Q1 2026. If you took that at face value you’d conclude Abu Dhabi prices almost doubled in six years. The reality is more subtle, and understanding it is the single most important thing any Abu Dhabi investor can learn.

The ADREC index, like most official residential indices worldwide, is computed as an average (or median) across whatever transactions happened in a given quarter. This works well when the composition of transactions is stable over time — but in Abu Dhabi, it has shifted dramatically. New luxury supply on Saadiyat, Fahid, Jubail, and Ramhan Islands has pulled the average price per transaction steadily upward, independent of what happened to any individual property. When a wave of AED 5M penthouses hits the market, the "average" price goes up — even though the same AED 1.5M apartment in Al Reem Island remained priced at AED 1.5M. This is known as composition bias.

To isolate true per-unit price change, Daar computes a repeat-sales price indexusing the methodology developed by Karl Case and Robert Shiller, now the standard for the S&P Case-Shiller indices in the United States. The method works as follows:

• For every property that has transacted at least twice, match its first sale with its most recent sale (matched by project name, property layout, and rounded area in square metres).
• Compute the per-unit annualised return between the two transactions, filtering out implausible returns (below -30% or above +50% annualised) and absurd holding periods (shorter than 6 months or longer than 10 years).
• For each quarter, take the median of all annualised returns from pairs whose second transaction falls in that quarter.
• Chain the medians into an index, rebased so Q1 2020 = 100.

The result, as of the latest data, is that the repeat-sales index stands at roughly 103.7 in Q1 2026 — just +3.7% above the 2020 baseline. That’s +4% over six years, not +96%. The gap between ADREC’s published index and Daar’s repeat-sales index is almost entirely attributable to new-build luxury supply entering the mix, not to existing units appreciating.

We build the index from 15,488 matched pairs as of the latest import, which is enough statistical power to produce a stable quarterly series. Both indices are shown side by side on the Dashboard so investors can decide which signal matters more to their use case. If you’re buying a specific existing apartment, the repeat-sales index is almost certainly closer to what you should expect. If you’re sizing the total value of the market for a capital-allocation decision, ADREC’s headline index might be what you want. Both are correct — they answer different questions.

Rental yield is the single most decision-relevant metric for a buy-to-let investor. It is also notoriously easy to compute badly. Daar’s yield methodology is built around two principles: match sales to rents at the narrowest possible granularity, and surface confidence transparently rather than pretending precision exists where it doesn’t.

Layout-level matching. For each district, we take all residential sales and rentals from 2025 onwards, group them by bedroom layout (studio, 1BR, 2BR, 3BR, 4BR+), and compute the median sale price and median annual rent for each layout that meets a minimum sample threshold (three sales, three rents). Gross yield for that layout is (annual rent median) / (sale price median). We then weight the layout yields by transaction volume to produce a district-level gross yield.

Pooled fallback.Some Abu Dhabi districts — particularly older inner-city neighbourhoods like Al Manhal, Al Mushrif, and Al Khalidiyah — have a large portion of transactions filed by ADREC under "unclassified" property layout. For these districts, per-layout matching would drop most of the sample. Instead, we pool all sales and rents district-wide and compute yield from the pooled medians. Districts using the pooled method are automatically tagged Low confidence, because we lose the layout-level control.

Net yield.From gross yield we subtract a flat 7% for operating expenses — typical for Abu Dhabi apartments across service charges (AED 10-25/sqft/year), property management (5-8% of rent), and a 5% vacancy allowance. This 7% assumption is conservative for larger units and freehold zones, slightly aggressive for leasehold communities. We publish the assumption explicitly so investors can re-underwrite at their own opex estimate.

Confidence tiers.Every yield figure is tagged High, Medium, or Low based on sample size. High = at least 100 sales and 30 rents in the district. Medium = 30 and 10. Low = below those floors, or pooled-fallback calculation. Low-confidence figures are still shown because they’re often the only signal available for emerging districts — but they should be treated as directional, not precise.

Listing adjustment. We do not use portal listing prices in the yield calculation. Listings systematically run 5-15% above transaction prices (the listing premium we track separately), and mixing them in would inflate apparent yields by 5-10%. Everything in the yield pipeline is transaction-grounded.

The Daar price forecast is a deliberately simple model: an ordinary least-squares regression of quarterly median price per square metre on quarter index (1, 2, 3…), fitted per district with a minimum of 8 quarters of history. We intentionally avoid ML, neural networks, or ensemble models because our back-testing (Section 1) shows that, at this volume of data, the simple regression is not materially outperformed by more complex alternatives — and simpler models are much easier to audit and explain.

The regression produces a point estimate for 1-year, 2-year, and 3-year horizons. Around each point estimate we attach a confidence band equal to plus-or-minus the empirical MAPE observed in walk-forward back-testing for that horizon, scaled by square-root-of-T (the standard statistical adjustment for error growth over longer forecast horizons). The result: a 1-year forecast typically shows a band of plus-or-minus 25-40% around the central estimate, widening to 55-80% at three years.

Those bands are intentionally wide. A tighter band would be prettier to display but dishonest — our back-tests show errors of exactly that magnitude historically, and there is no credible reason to believe the future will be more predictable than the past. If a forecast says "AED 2M base", the realistic range is AED 1.4M to AED 2.6M at 1 year, AED 0.9M to AED 3.1M at 3 years.

Confidence label. Each forecast gets a label (High, Medium, Low) based on two factors: the number of quarters available for the regression (20+ = high, 12-19 = medium, 8-11 = low) and the R-squared of the fit (0.5+ = full credit to the history). Districts with short histories or noisy quarterly series get Low confidence automatically.

What the model doesn’t capture.Policy changes (new Golden Visa tiers, mortgage regulations), supply shocks (a single mega-project delivering 3,000 units), and global macro shifts (interest rates, oil prices, geopolitical events) are all outside the model’s scope. The forecast is a trend extrapolation, not a causal prediction. Use it as one input among several, not as the final answer.