Asset Mechanics

RCF-AI Land Exploitation Risk

Context-Based Risk Profiles and Buffer Assessment

Card version: 1.51
Draft date: 24 October 2026
System release: RCF-AI 2026.01
Publication date: 2026-10-24 12:54:18
Status: Reviewed and accepted

At a Glance

This card explains the context-based risk profiles used to assess land-development (grondexploitatie) buffers. It is intended for municipal finance teams, land and project developers, analysts and decision-makers.

Decision role: second-opinion and validation support. The responsible organisation decides on reserves and risk appetite.
Evaluation scope: Land price (grondprijs, ground quote 0.35); Netherlands; annual horizons 1–10 years; VaR15 (15th percentile, lower tail); evaluation runs of 24–26 September 2026.

Main findings:

1. System Overview and Intended Use

The tool combines context-based risk profiles with project exposures, phasing and context-based correlations to assess project and portfolio downside. The dashboard shows buffers against both the current budget and the modelled expected outcome. End-of-project value (eindwaarde) may be used as the cash-flow reporting basis.

Economic context → context-based risk profiles and correlations → exposures and timing → project and portfolio downside → buffer need.

The dashboard enables users to inspect the basis of an estimate:

Features and comparable years can change with the factor and current context. They explain predictive relevance, rather than establishing causation. Section 5 describes the limits of this evidence.

Other outputs: aggregated values at requested percentiles:

• result
• present or end value
• contribution of risk per driver

2. Evaluation Approach and Evidence Base

2.1 Model Selection and Validation Requirements

We continuously search for, and further develop, financial, statistical and context-based machine-learning methods through rigorous out-of-sample validation and benchmarking against our high validation standards as described in our Validation policy.

Minimum standards for selectable models can be found on:

• https://assetmechanics.org/insights/trustworthy-ai/ and
• https://assetmechanics.org/insights/out-of-sample-validation/.

Although the exact underlying method may differ per risk factor, the main methodology is guided by the framework described under:

• https://assetmechanics.org/insights/predictive-rcr/.

We call each selected method the context-based model.

Reported results apply to the latest selected best model for that factor at the stated release and assessment date. “Best” means the best eligible candidate under our criteria for quantile loss, coverage error and performance across horizons.

Selection criteria: candidates are compared with BASE on the same out-of-sample VaR forecast per horizon, on coverage error and quantile (pinball) loss at the target percentile. We further look into independence of breaches (share of independence tests rejecting at 5%) per horizon and calculate the mean over horizons 1–10.

Our validation requirements are to:

• Test the complete modelling and selection process on unseen outcomes, using information available at the forecast date.
• Match the model and benchmark on observations, target definition, horizon, quantile and metric scaling.
• Distinguish primary claim evidence from diagnostics and planned tests, retain reproducible records, and report material uncertainty and context limitations.
• Support user checks of inputs, drivers and historical support, with reassessment after material changes.
• Ensure that the model methodology provides empirical reference class evidence and key drivers.

2.2 Methods and Data

Validation design: walk-forward (expanding window) on the last 20 realised forecasts per horizon is the primary evaluation reported in Chapter 3; time-ordered cross-validation is used within candidate model hyperparameter optimization and selection is based on data before the evaluation window.
Executed design and review status: executed 22 September – 2 October 2026 (runs registered in the model registry); review status Reviewed and accepted.

Cross-validation, where used, evaluates held-out subsets under the specified split scheme. Its role in selection or primary evaluation is rolling-origin, time-ordered folds before the evaluation window; used for hyperparameter and model/feature selection only, not for the reported results in Chapter 3.

Data basis: annual Netherlands series; cumulative percentage change in land price over h years from the forecast origin (dataset as of 22 May 2026; realised fwd = 1 outcomes 1914–2024, i.e. 111 years in total). Evaluation covers the last 20 realised values per fwd. Source attribution: CBS, KEF, DNB.

Data limitations:

• Some features are only updated yearly, so real-time availability and revisions throughout the year are limited.
• Outcomes of consecutive years overlap for fwd ≥ 2, so the 20 evaluation outcomes per horizon are not independent.
• One breach moves coverage by 5.0 pp.

2.3 BASE

BASE is the relatively static reference using the full eligible historical record without target-specific contextual selection. Its history expands over time; only outcomes available at a forecast origin are eligible.

History and updates: from 1914 onwards; expanding window recalculated for every forecast using only outcomes realised before forecast period (years ≤ t − h). For fwd = 10 the last evaluated reference date is 2015; for fwd = 1 it is 2024.
Comparison configuration: RCF-AI 2026.01; the model and BASE are scored on identical origins and realised outcomes (verified match), 20 per horizon, from 2005–2024 (fwd = 1) to 1996–2015 (fwd = 10); VaR15 only, no other quantiles evaluated.

3. Performance and Interpretation

The following block is completed for the relevant risk factor and quantile. It reports the model and BASE on matched observations for each annual horizon.

Risk factor: Land price (grondprijs, Netherlands).
Validation design / result set: walk-forward, expanding window, 20 data points within test set per horizon; context-based run 20260924T141622_8495fd vs BASE run 20260926T121127_df6b46.
Tail, quantile and target protection: lower tail (price decline); 15th percentile (VaR15); target coverage 85% of outcomes at or above the boundary.
Outcome basis: cumulative percentage change in land price over h years (fwd) from the forecast reference date.

For evaluation we use standard percentile metrics absolute coverage error and quantile loss. For both metrics, positive improvement is better and negative improvement is deterioration. A zero BASE denominator gives N/A; report the raw values and absolute change. Near-zero denominators can make relative improvements unstable. A percentage-point change is different from relative percentage improvement.

3.1 Coverage Error

Coverage is the share of outcomes on the protected side of the predicted boundary.

Signed coverage error (pp) = observed coverage (%) − target coverage (%).

An average absolute coverage error of zero is the target. The average signed coverage is provided as a supportive diagnostic. Negative values indicate undercoverage; positive values indicate overcoverage. Improvement is measured by the reduction in the error’s magnitude:

Improvement (%) = 100 × (absolute BASE error − absolute model error) / absolute BASE error.

Interpretation:

3.2 Quantile Loss

Quantile loss measures the distance between predicted quantiles and realised outcomes, with asymmetric penalties determined by the quantile. Lower loss is better.

Percentage definition: mean pinball loss at α = 0.15 of the predicted 15th-percentile boundary q against the realised cumulative change y: α·(y − q) if y ≥ q, else (1 − α)·(q − y); y and q are expressed as fractions and the loss is multiplied by 100, so the unit is percentage points of cumulative change; simple mean over the 20 origins per horizon.

Improvement (%) = 100 × (BASE loss − model loss) / BASE loss.

* | ** | *** improvement significant at the 10% | 5% | 1% level (i.e. at least 1.65 | 1.96 | 2.58 × the standard deviation of the loss differences / √20, as % of BASE loss).

Interpretation:

3.3 Quantile Estimation Bias

Quantile Estimation Bias measures whether the predicted VaR boundary lies, on average over the test period, above or below the realised percentile in that test period.

It is signed and provided for supporting diagnostics; it is not used as main evaluation metric.

For the lower-tail VaR reported here, a negative bias means the boundary is on average below the realised 15th percentile (conservative); a positive bias means it is above it (not conservative enough).

Quantile Estimation Bias (pp) = mean of the predicted 15th-percentile boundary q over the 20 test origins − 15th percentile of the realised cumulative changes y over the same origins.

y and q are expressed as fractions and the bias is multiplied by 100, so the unit is percentage points of cumulative change; the realised percentile interpolates linearly between the ordered outcomes. Improvement is measured by the reduction in the bias’s magnitude:

Improvement (%) = 100 × (absolute BASE quantile bias − absolute model quantile bias) / absolute BASE quantile bias.

3.4 Supporting Checks and Client Material

We provide clients with additional robustness-validation material:

• historical out-of-sample charts comparing central predictions and downside boundaries with realised outcomes, including expected and observed breaches where applicable.
• complete out-of-sample buffer backtest over 20 years

Each report identifies its factor, horizon, period, validation design and material limitations.

Available client material:

• main predictive drivers per prediction;
• reference class evidence per prediction containing the predictively most comparable years;
• internal out-of-sample charts of realised outcomes against the 15th-percentile boundary per test year exist for land price, horizons 1–12, over the 20-year evaluation windows (alternative models comparison report).

4. Validated Scope and Responsible Use

Before a reserve decision, the responsible professional checks:

1. Inputs are current, correct and fully mapped; the important drivers are complete and appropriate to the case.
2. Profiles, horizons, percentiles, reference outcomes, timing and portfolio assumptions fit the decision.
3. Dashboard features, comparable years, SE indicators and warnings have been reviewed, and material exceptions resolved.

Stop / escalation route: Conditions requiring review:
- input data older than one year requires recalculation of the context dependent risk model with most actual context;
- extreme shocks in the macro-economy since last cut-off date require analysis against available reference class evidence.

The tool does not authorise automatic reserve changes, transactions or binding development decisions. Accountable human review and financial, legal and planning due diligence remain necessary; review alone does not extend validated scope.

5. Known Limitations and Uncertainty

• Limited tail evidence. For some rare adverse outcomes, relatively small contextual samples and overlapping long horizons may limit calibration conclusions.
• Changing conditions. The validation evidence is represenative for the model trained on 90 years of macro-economic data and tested out-of-sample over the last 20-30 years. Historical relationships and comparable years may not represent a new regime that would fall outside the distribution of this dataset and may lead to fallback to a broader history.
• Transfer and aggregation. Evidence for one factor or horizon does not validate another or establish portfolio coverage. Dependence, shared exposures and timing matter.
• Conditional buffers. Inputs and risk settings affect requirements. Better profile metrics do not imply an equal percentage reduction in reserves. Also the other risk factors need to be considered.

Additional release limitations: the context-based model's breaches are clustered at horizons 3–4 and 8–9, so breaches may arrive in consecutive years rather than spread out.

Model review process: the breach dependence and incoherence between subsequent years are topics that are further investigated and addressed in our continous model improvement process. To increase precision (standard errors) of the validation metrics we plan to test longer holdout periods in future validation runs.

6. Data Use and Transparency

Data handling: CBS, KEF and DNB are open data sources; customer data is not shared with other customers or parties; person related land data is not loaded and processed; data and usage is encrypted.
External AI: not used.

The card reports results and material limitations. Users receive case-specific explanations needed to check responsible use. Model identities, architectures, code, proprietary selection rules and confidential customer records remain internal.

7. Version History and Updates

Review arrangements: Asset Mechanics, September 2027 as part of the yearly model review cycle. Reassessment also follows material changes to data, models, features, benchmarks, scope or reliance. A replacement model requires updated evidence and release-linked metrics.

8. Assurance and Validation Roadmap

Validation review: context-based run 20260924T141622_8495fd vs BASE run 20260926T121127_df6b46 (model registry and alternative models comparison report).
Sign-off: reviewed and accepted by CTO & CEO.

9. Evaluation Record and Contact

Evaluation record: land price: context-based run 20260924T141622_8495fd vs BASE run 20260926T121127_df6b46; alternative models comparison report (system_card).
Evaluation owner: Asset Mechanics R&D.
Public contact: [email protected].

Basis: requirements v1.1, public structure v1.8 and the Core Validation Framework.