Loading...
Loading...
Oliva research
Every probabilistic or heuristic reading on the platform links to the section below that defines its source data and computation rule.
Oliva produces an independent 0–100 quality rating for every Dubai off-plan project — the **Oliva Score**. It is the same number for every investor and cannot be paid for or influenced by developers. This page explains exactly what the score measures, where the underlying data comes from, how it is computed, and where its limits are.
We publish this because a score is only as trustworthy as the method behind it. Everything below reflects the scoring engine (`oliva-1.0.0`) that runs in production today.
---
The Oliva Score is a **weighted average of six differentiating dimensions**, each itself a confidence-weighted blend of normalized metrics drawn from authoritative Dubai and global data sources. The engine scores every published listing on a 0–100 scale and aggregates listing scores up to a project (promotion) score.
Two design choices shape the whole system:
1. **We only weight what differentiates.** A dimension or metric that is flat across the Dubai market — i.e. it does not separate a good project from a bad one — is given little or no weight, even if it sounds important. Macro context is the clearest example (see below). 2. **Missing data is dropped, not penalized.** If a metric has no reliable value for a given project, it is removed from that dimension's weighted average and the remaining weights are renormalized — rather than scoring the gap as zero. A project is never punished for a data point Oliva simply does not have yet.
---
The overall score is driven by the six dimensions below. The weights are the live composite weights used by the engine.
| # | Dimension | Weight | What it measures | |---|-----------|:------:|------------------| | 1 | **Financial Value** | **22%** | Whether the price is attractive relative to the area, the income the property can generate (gross and net rental yield, cap rate), and where it sits on the DLD price percentile. The single most predictive driver of long-term return — over-paying in a great location still erases years of gains. | | 2 | **Market Dynamics** | **20%** | The health and momentum of the local market: transaction velocity and volume, absorption rate, off-plan vs resale mix, official DLD sale-price index trend, and multi-year area price CAGR (5/10/20-year). | | 3 | **Location Quality** | **20%** | The physical surroundings that cannot be changed after purchase: amenity density, transit and metro accessibility, proximity to schools and healthcare, walkability, and sub-area premium. | | 4 | **Developer Trust** | **15%** | The reliability of the developer behind the project: lifetime delivery history, years active, project and unit count, resale premium on completed projects, and regulatory/licensing standing. In off-plan you are buying a promise — this dimension quantifies how likely it is to be kept. | | 5 | **Risk Assessment** | **15%** | The downside: price volatility, oversupply risk, concentration of a single developer or unit type, maximum historical area drawdown, and construction-completion risk. Understanding risk matters as much as understanding return. | | 6 | **Liquidity & Exit** | **8%** | How easily and quickly an investor can get out: resale transaction frequency, area transaction velocity, days on market, rental demand, and an estimated exit timeline. The best investment is worthless if you cannot exit when you need to. |
**The six weights sum to 100%.**
Oliva also computes a seventh dimension, **Macro Context** (UAE GDP growth, interest-rate environment, inflation, currency stability, oil-price stability, and global volatility indices). It is calculated nightly and surfaced separately in the product as **Market Context**, but it carries **0% weight in the overall composite score**.
The reason is deliberate and honest: macro factors move *every* Dubai project in a given quarter in essentially the same direction. A rate cut lifts all projects; a spike in global volatility weighs on all of them. Because macro does not *differentiate* one project from another, including it in the composite would add noise without improving the relative ranking that investors actually use the score for. We therefore show it as context rather than baking it into the headline number.
An earlier version of the engine also scored a **Regulatory & Legal** dimension (DLD registration, escrow compliance, Trakheesi permit, Oqood registration, insurance). It was removed from the composite because, in production, virtually every listed project cleared these checks — the metrics were effectively flat (standard deviation < 3) and so did not separate projects. Regulatory standing is still verified operationally during onboarding; it simply is not a *scoring* lever because it does not discriminate between the projects on the platform.
---
The pipeline runs in four conceptual steps. We describe each qualitatively; we do not publish exact coefficients or cutoffs, because those are continuously recalibrated against the market.
Raw inputs live on wildly different scales — a price per square foot, a rental yield percentage, a distance in metres, a transaction count. They cannot be combined directly. Each metric is mapped onto a common 0–100 footing using the shape that fits its economics:
The floors, ceilings, and targets are calibrated against observed Dubai distributions and reviewed when the market shifts.
Newer and more complete data counts for more. Each metric carries a data source and an implicit confidence based on freshness and coverage. Where a project-level signal is too thin to trust, the engine falls back to the area-level equivalent (for example, observed rental yield falls back from project-specific Ejari contracts to the area average when there are fewer than five contracts to learn from). Stale market data is down-weighted relative to recent data.
Within each dimension, the normalized, confidence-weighted metrics are combined into a single 0–100 dimension score. Metrics with no reliable value are dropped and the dimension's remaining metric weights are renormalized, so a missing input neither inflates nor deflates the dimension.
The six dimension scores are combined into the overall 0–100 Oliva Score using the dimension weights in the table above. If an entire dimension has no usable signal, it too is dropped and the surviving dimension weights are renormalized — the composite is always a weighted average over the dimensions that actually have data, never a partial sum padded with zeros. A data-completeness percentage is recorded alongside every score so coverage is transparent.
---
Alongside the Oliva Score, Oliva estimates a **fair price per square foot** for a unit and compares it to the asking price. This answers a different, complementary question: *"Is this specific listing priced fairly versus what the market actually pays for comparable units?"*
Fair price is produced by a **hedonic log-PSF regression** fitted nightly on Dubai Land Department transactions. In plain terms, a hedonic model learns how much each attribute of a property contributes to its price, then uses those learned contributions to predict a fair price for any given unit.
The model is a **two-stage fit** run nightly on DLD transaction history:
For a given listing, the predicted `log(fair PSF)` sums:
That last point is an important honesty caveat. Some attributes that clearly affect price — finishing quality, parking count, ceiling height, service charge, off-plan handover timing, and post-handover payment plans — **do not appear in DLD transaction records at all**, so they cannot be learned from history. They are applied at prediction time as configured percentage **priors** (multipliers) rather than fitted coefficients. As Oliva accumulates a long enough record of its own fair-price predictions versus realised prices, these priors will be replaced by values learned from data.
Every fair-price prediction comes with a **95% confidence interval** derived from the model's residual variance plus the uncertainty of the building-level estimate, and with a **comp tier (1–5)** that grades how much real comparable evidence backed it — from Tier 1 (five or more recent same-building comps, high confidence) down to Tier 5 (no developer history; area-and-bedroom baseline only). The tier and interval are surfaced so that a fair-price estimate is never presented as more certain than the evidence supports.
---
Every data point that feeds the Oliva Score or the fair-price model traces to an authoritative source:
---
We would rather state the limits plainly than overstate the score's precision:
---
*This document describes the production scoring engine (`oliva-1.0.0`). Dimension weights and calibration are reviewed as the Dubai market evolves; the score is recomputed on a regular cadence as new transactions, rent contracts, and regulatory data arrive.*
Worked examples and comparative analysis from the Oliva research team.