Bytero Engine™
Our Machine Learning XGBoost algorithm built on 185 property attributes with visual recognition AI.
Bytero Engine sets the market price for the entire real estate market. XGBoost model over 185 property attributes with visual recognition AI over photos, calibrated per region and segment, holdout-validated and continuously retrained. Deal screening, due diligence and portfolio monitoring for funds, asset managers and valuation teams.
Prices, rents and values
Ten values for every address: what the market sees and what the law requires. All of them are available through the API.
The estimate of what the property is actually worth: model prediction calibrated on real prices from closed transactions, not on listings.
The current asking level: what comparable properties are listed for right now in the same location and segment.
Achievable monthly rent from a second, independent model trained on the rental market.
Gross annual rental yield: predicted rent times 12, divided by the predicted price, consistent across the whole market.
Price per square metre of usable area, comparable across layouts and cities.
Daily price snapshots reveal drops, increases and how long a listing sits on the market.
The cost of rebuilding the structure under decree 492/2004, the base for technical value and insurance.
Reconstruction value reduced by wear based on age and condition, computed the appraiser way.
The expert value: technical value adjusted by the location coefficient to market conditions.
The correct base for an insurance contract, derived from reconstruction value. Protects against under- and over-insurance.
Architecture and training process
The full market
Bytero Engine processes the full market in real time. Every property is tracked continuously: attributes, photos, price movements and conversion history. The inputs also include real sale prices from closed transactions, which anchor the model in prices actually achieved. This data is the foundation of both training and calibration - without it no prediction would reflect real market dynamics. Bytero is the market data layer, not a one-off snapshot.
Cleaning, deduplication and outlier filtering
Raw data passes through a multi-stage pipeline. Perceptual-hash image matching combined with spatial and geometric cross-checks collapses the same apartment across the entire market. We canonically normalize disposition, condition, construction type and energy rating, and unify currency. For training the model weights price records against dominance by larger segments and iteratively filters extreme price deviations - so Bytero Engine learns typical market behavior, not agent anomalies or dataset outliers.
Feature engineering and visual analysis
From each property we generate dozens of features across five categories: location (distances to city centre, transport, schools, supermarkets), geometry (area, building age, floor), amenities (elevator, parking, balcony, garage), geographic encoding of regions and districts, and a visual-financial interior score. The visual score comes from a multi-modal computer-vision model that rates the apartment from its photos. For the prediction every feature is a first-class input.
Two independent XGBoost models - sale and rent
We train two independent XGBoost regressors - one for sale price, one for monthly rent. Each is an internal ensemble of several variants optimized for different scales and robustness against skewed price distributions. XGBoost (gradient-boosted trees) is the most accurate class of algorithms for tabular data, and Bytero Engine tunes it for the real estate context with custom regularization, validation and an anti-overfitting strategy. The output is a single market price - not a compromise between models.
Sale-price and regional calibration
Sellers price optimistically and the gap to what buyers actually pay varies by segment. After training, Bytero Engine measures the median residual between the model and the observed market on holdout data, per segment, and applies a calibration offset to every prediction. The same mechanism captures regional specificity - capital districts, resort districts and rural counties operate under different rules and receive their own offsets. Resort pricing is therefore not collapsed into capital-region dynamics.
Holdout validation and continuous retraining
Every new model candidate is validated on a holdout set it has never seen - we measure R², MAE, MAPE, MedAPE, the share of predictions within specific bands, and a per-segment scoreboard. Only models that beat the current champion go live on every key metric. After every market refresh Bytero Engine backfills prices for new properties and at regular intervals retrains on the latest data. No stale models - always current calibration.
The market price
For each property Bytero Engine sets four prices: the market price (the level of comparable properties on the market) and the sale price (the level at which transactions actually close) on the sale side, plus gross and net monthly rent on the rental side. From them it derives the yield and the deviation of the current price from the market price. The sale price is derived from the market price through a calibration layer that accounts for the region, current mortgage rates (ECB MIR), time on market and the property's deviation from comparable properties. The output is structured, dated and versioned: ready for decision-grade use in deal screening, portfolio monitoring and valuation reports. Bytero Engine sets the price that reflects the real market - with no manual comparison and no compromise between models.
Visual analysis of the interior
A computer-vision model processes the property photos and quantifies attributes that a broker would otherwise assess during a physical viewing.
Multi-modal LLM
Bytero Engine uses a multi-modal computer-vision model that processes all property photos at once and returns a structured assessment. It adds a dimension pure tabular data cannot capture: what the property actually looks like. The same apartment with identical area and location may be move-in ready or require a full renovation - and the model sees the difference.
Nine scored attributes
On a 0-10 scale we rate nine attributes: overall condition, kitchen, bathroom, furniture, floor, windows, natural light, photo quality and renovation need. The score distribution is calibrated so the upper extreme covers only a small share of the portfolio - the model discriminates, it does not regress to the mean. Every attribute is a separate input to the prediction pipeline.
Impact on the prediction
Two apartments with identical area, location and disposition can have market values that differ by tens of thousands of euros - depending on the state of the interior. The visual score enters the prediction as a full-featured input and explains the variance a purely numeric model would miss. Without it Bytero Engine would conflate a renovated apartment with one in original state at the same layout.
Versioned scoring
On every property you see the aggregated visual score together with the version of the scoring framework. With each model upgrade we can compare retroactively which version delivered which improvement, and apply a new revision retroactively. No black box - every score has an auditable origin, evaluation date and reviewable metadata.