Geospatial Analytics, Modelling and Decision Support

Introduction and Context

Urban and regional planning increasingly depends on evidence that is spatial, dynamic and multidimensional. Questions such as where to locate housing, how to protect agricultural land, which transport corridor will best serve future growth, and how climate hazards may affect infrastructure cannot be answered adequately with maps alone. They require the integration of geographic information systems (GIS), remote sensing, spatial statistics, suitability analysis, scenario planning, urban modelling and decision-support methods.

Geospatial analytics converts location-referenced data into interpretable evidence. It combines spatial data management, measurement, statistical inference, simulation and visual communication. The purpose is not to replace professional judgement or democratic deliberation, but to make assumptions explicit, compare alternatives systematically and reveal spatial consequences.

At postgraduate level, the planner must understand both the technical workflow and its limitations. A technically precise model can still produce poor decisions if the data are biased, the spatial unit is inappropriate, the weights are arbitrary or the results are communicated as predictions rather than conditional estimates.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the role of GIS, remote sensing and spatial databases in planning.

2. Construct and interpret spatial weights matrices and common spatial-statistical tests.

3. Perform suitability analysis using constraint mapping and weighted criteria.

4. Apply multicriteria decision analysis (MCDA), including normalization, weighting and sensitivity testing.

5. Develop and compare land-use and infrastructure scenarios.

6. Distinguish descriptive, explanatory and simulation-based urban models.

7. Evaluate uncertainty, data quality, ethics and governance in geospatial decision support.

8. Communicate analytical results responsibly to technical and non-technical audiences.

1. Geospatial Evidence as a Planning System

A planning geospatial system consists of five linked elements:

• Data: geometry, attributes, time stamps, imagery, survey observations and administrative records.
• Methods: overlay, proximity analysis, classification, interpolation, spatial statistics and simulation.
• Models: representations of relationships among land use, movement, environment, population and infrastructure.
• Decisions: policies, investments, regulations or site selections.
• Institutions: agencies, communities, laws and accountability arrangements.

A GIS stores vector data (points, lines and polygons) and raster data (regular cells). Vector data are appropriate for parcels, roads and administrative boundaries; raster data are especially useful for elevation, land cover, temperature and suitability surfaces. A coordinate reference system must be selected consistently. Mixing geographic coordinates in degrees with projected coordinates in metres can invalidate distances, areas and buffers. For local engineering and planning analysis, an appropriate projected coordinate system is normally preferable.

A relational spatial database should preserve geometry, attributes, source, date, positional accuracy, processing history and access rights. Metadata are not optional documentation: they determine whether a dataset is fit for a particular decision.

Remote sensing

Remote sensing provides repeated observations of the Earth's surface from satellites, aircraft or unmanned aerial systems. Interpretation depends on spatial, spectral, temporal and radiometric resolution. A land-cover classification is not identical to a land-use classification: a spectral signature may identify vegetation, whereas land use concerns the social or functional purpose of land.

For two spectral bands, a normalized difference index may be written as

\[

NDI=\frac{R_2-R_1}{R_2+R_1},

\]

where \(R_1\) and \(R_2\) are atmospherically corrected reflectances. The index is dimensionless and is meaningful only when the selected bands, preprocessing and interpretation are appropriate. Vegetation indices commonly use near-infrared and red reflectance, but threshold values are sensor-, season- and context-dependent. Classification accuracy should be assessed using independent reference samples, a confusion matrix and measures such as producer's and user's accuracy; overall accuracy alone may conceal poor performance for minority classes.

2. Spatial Analysis and Spatial Statistics

Spatial data violate assumptions that are often reasonable for ordinary tabular data. Nearby observations may be more similar than distant observations, and observations may vary across space. These properties are known as spatial dependence and spatial heterogeneity.

Spatial weights matrices

A spatial weights matrix \(W=[w_{ij}]\) represents the neighbourhood structure. Usually \(w_{ii}=0\), because a location is not its own neighbour. Common definitions include:

• Contiguity: polygons share an edge (rook) or edge/corner (queen).
• Distance band: locations within a specified distance are neighbours.
• k-nearest neighbours: each location is linked to its k closest locations.
• Distance decay: influence declines with distance, for example \(w_{ij}=d_{ij}^{-p}\) when \(d_{ij}>0\).

Weights may be binary or continuous and are commonly row-standardized:

\[

w^*_{ij}=\frac{w_{ij}}{\sum_j w_{ij}},

\]

provided the row sum is non-zero. Row standardization makes each row sum to one, so the spatial lag \((W^*y)_i\) is a weighted local average. It does not make the matrix symmetric, and it changes the interpretation of statistics. Isolated locations require an explicit treatment, such as a larger distance threshold or k-nearest-neighbour structure; silently leaving zero rows can create misleading results.

Weights are a modelling choice, not a natural fact. Results should be tested under plausible alternatives, such as queen versus rook contiguity or different distance bands.

Global Moran's I

For observations \(x_i\), mean \(\bar{x}\), deviations \(z_i=x_i-\bar{x}\), and weights \(w_{ij}\), global Moran's I is

\[

I=\frac{n}{S_0}\frac{\sum_i\sum_j w_{ij}z_i z_j}{\sum_i z_i^2},

\qquad S_0=\sum_i\sum_j w_{ij}.

\]

For a randomization or random-permutation null hypothesis, the commonly used expected value is \(E(I)=-1/(n-1)\). A positive value greater than the expectation indicates spatial clustering of similar values; a value near the expectation is consistent with spatial randomness; a negative value indicates neighbouring dissimilarity. Statistical significance requires a variance or permutation procedure. A positive statistic alone is not proof of significant clustering.

Procedural interpretation of a significance test

A defensible workflow is:

1. Define the variable, spatial units, time period and substantive null hypothesis.

2. Inspect the distribution and map the values; identify missing data and extreme values.

3. Specify and justify \(W\), including contiguity, distance or k-nearest-neighbour rules.

4. Calculate the observed statistic \(I_{obs}\).

5. Generate many random permutations of the observed values among the fixed locations, recalculating Moran's I each time.

6. Estimate a two-sided permutation p-value as

\[

p=\frac{1+\#\{|I_b-E(I)|\ge |I_{obs}-E(I)|\}}{1+B},

\]

where \(B\) is the number of permutations. The plus-one correction avoids a zero estimated p-value.

7. If the p-value is below the preselected significance level, reject spatial randomness for the specified weights and variable. Otherwise, do not claim that clustering is absent; state that evidence is insufficient.

8. Repeat the analysis using plausible weights and, where appropriate, alternative spatial units.

The test concerns a pattern under a particular null model. It does not establish causation, nor does it identify the mechanism producing the pattern.

Local indicators and spatial regression

Local Moran statistics help identify high-high, low-low, high-low and low-high local associations. They require correction or cautious interpretation when many locations are tested, because multiple testing increases false positives.

If residuals from an ordinary least squares model are spatially autocorrelated, coefficient estimates may still be unbiased under restrictive conditions, but standard errors and inference can be unreliable. Spatial lag and spatial error models represent different mechanisms:

\[

y=\rho Wy+X\beta+\varepsilon

\]

models dependence in the dependent variable, while

\[

y=X\beta+u,\qquad u=\lambda Wu+\varepsilon

\]

models spatial dependence in the error process. These models require careful specification, identification and diagnostics. A significant spatial coefficient is not automatically evidence that a policy variable causes the observed outcome.

3. Suitability Analysis and MCDA

Suitability analysis identifies locations that meet constraints and perform well on selected criteria. It normally includes four stages:

1. define the decision objective and feasible area;

2. apply exclusion constraints;

3. standardize eligible criteria to a common preference scale;

4. aggregate criteria and test sensitivity.

Constraints are usually Boolean masks, such as excluding protected wetlands, unstable slopes or land outside a service boundary. Criteria express degrees of preference, such as proximity to transit, low flood exposure or land cost. A constraint should not be disguised as a low score: if development is legally prohibited, it should be excluded or explicitly modelled as a hard restriction.

For alternatives \(i=1,\ldots,m\) and criteria \(j=1,\ldots,k\), a simple additive model is

\[

S_i=\sum_{j=1}^{k} w_j r_{ij},

\qquad \sum_jw_j=1,

\]

where \(r_{ij}\) is a standardized score, usually between 0 and 1, and \(w_j\) is a non-negative weight. Additive aggregation assumes criteria are sufficiently compensatory: a weakness on one criterion can be offset by strength on another. If that assumption is unacceptable, use thresholds, veto rules, outranking or a non-compensatory method.

For a benefit criterion, min-max standardization is

\[

r_{ij}=\frac{x_{ij}-\min(x_j)}{\max(x_j)-\min(x_j)}.

\]

For a cost criterion,

\[

r_{ij}=\frac{\max(x_j)-x_{ij}}{\max(x_j)-\min(x_j)}.

\]

These formulas require a non-zero range and are sensitive to outliers. The reference range and direction must be documented.

Worked MCDA example

Suppose three candidate sites are assessed using normalized scores for access, environmental suitability and cost. The weights are 0.40, 0.35 and 0.25 respectively.

| Site | Access | Environment | Cost | Weighted score |

|---|---:|---:|---:|---:|

| A | 0.80 | 0.60 | 0.70 | \(0.40(0.80)+0.35(0.60)+0.25(0.70)=0.705\) |

| B | 0.65 | 0.85 | 0.60 | \(0.40(0.65)+0.35(0.85)+0.25(0.60)=0.7175\) |

| C | 0.90 | 0.50 | 0.40 | \(0.40(0.90)+0.35(0.50)+0.25(0.40)=0.635\) |

Thus B ranks first, A second and C third. Rounded to three decimal places, B's score is 0.718, not 0.667; its margin over A is 0.0125, or approximately 0.013. The small margin indicates that the ranking should not be treated as decisive without sensitivity analysis.

If the access weight rises from 0.40 to 0.50 and the environmental weight falls to 0.25, with cost unchanged, the new scores are A = 0.715, B = 0.7075 and C = 0.695. A then ranks first. This demonstrates weight sensitivity and the importance of reporting alternatives rather than presenting one ranking as objective truth.

Weight elicitation and robustness

Weights may be obtained through expert workshops, pairwise comparison, budget or policy targets, stakeholder surveys, entropy methods or participatory negotiation. No method removes value judgements. A transparent report should state who supplied the weights, how disagreements were handled, and whether criteria overlap. Correlated criteria can double-count a characteristic, such as using both distance to a road and travel time without examining their relationship.

Robustness analysis varies weights, thresholds, data and classification rules. Useful outputs include rank-change tables, tornado diagrams, scenario envelopes and maps of locations that remain selected across many plausible assumptions.

4. Scenario Planning and Urban Models

A forecast asks what is likely under stated assumptions; a scenario asks what could happen under a coherent set of assumptions. Scenario planning is appropriate when demographic, economic, technological or climatic uncertainty is substantial.

A useful scenario process is:

1. establish the baseline and current trends;

2. identify driving forces and critical uncertainties;

3. construct internally coherent narratives;

4. translate narratives into spatial and quantitative assumptions;

5. model land use, transport, infrastructure and environmental effects;

6. compare indicators and distributional consequences;

7. test policies under multiple futures.

Typical scenarios include trend continuation, compact growth, dispersed growth, transit-oriented development, climate adaptation or constrained-resource futures. Scenarios should not be confused with predictions. Their value lies in stress-testing strategies and identifying actions that perform acceptably across futures.

Urban models

Urban models range from simple indicators to highly complex simulations:

• Gravity and spatial-interaction models estimate flows between origins and destinations. A generic form is \(T_{ij}=K O_iD_j f(c_{ij})\), where \(O_i\) and \(D_j\) represent origin and destination effects and \(f(c_{ij})\) is a calibrated impedance function. The constant and functional form depend on the model specification; the equation is not a universal law.
• Land-use–transport interaction models represent feedback between accessibility, development and travel. They require calibration and may be sensitive to zoning, network assumptions and behavioural parameters.
• Cellular automata represent transitions among neighbouring cells according to rules. They are useful for exploring spatial growth but can produce realistic-looking patterns without proving realistic causal mechanisms.
• Agent-based models represent heterogeneous actors and rules. They can examine household, developer or traveller behaviour, but parameterization and validation are demanding.
• System-dynamics models use stocks, flows and feedback loops to study aggregate change over time.

Model validation should distinguish calibration, verification and validation. Calibration estimates parameters using data; verification checks whether the implementation follows the specification; validation assesses performance against independent observations or known conditions. A model can reproduce past patterns yet fail under changed policy conditions because structural relationships are not stable.

5. Decision-Support Workflow and Communication

A professional workflow is iterative rather than linear:

1. frame the decision and identify affected groups;

2. define spatial and temporal boundaries;

3. assemble, document and quality-check data;

4. select methods consistent with the decision and data;

5. construct constraints, criteria, weights and scenarios;

6. run diagnostics and sensitivity tests;

7. compare alternatives using transparent indicators;

8. communicate uncertainty and trade-offs;

9. obtain review and revise the analysis;

10. monitor outcomes after implementation.

Good cartography uses an appropriate scale, legible symbology, clear legend, north arrow where useful, scale information, source, date and uncertainty statement. Choropleth maps should generally use normalized rates or ratios rather than raw counts when polygon populations differ. Classification affects interpretation, so equal interval, quantile and natural-break choices should be justified.

Dashboards can improve access but may encourage over-simplification. A decision-support product should provide downloadable data, definitions, processing notes, scenario assumptions and a record of model version. Reproducible scripts and version control reduce hidden changes and make peer review possible.

6. Illustrative Integrated Case: Growth and Flood Risk in the Riverbend Region

Background

Riverbend is an illustrative metropolitan region of 850,000 people. Population growth is concentrated along a river corridor containing wetlands, a railway, an existing bus network and several low-income neighbourhoods. The regional authority must identify locations for 18,000 additional dwellings over fifteen years while reducing flood exposure and avoiding unnecessary agricultural conversion.

Problem

A previous plan used a single development map and treated all vacant land as equivalent. It did not account for flood depth, access to public transport, network capacity, social vulnerability or cumulative wetland effects.

Analysis and approach

The planning team assembled parcel data, a digital elevation model, flood-depth simulations, land cover imagery, transit stops, road capacity, sewer-service areas, agricultural soils and census indicators. All layers were projected into a common metric coordinate system and assigned metadata.

Wetlands, the active floodway, protected habitat and land outside the sewer-service expansion boundary were initially treated as constraints. Remaining parcels were evaluated using five criteria: transit accessibility, flood safety, proximity to existing utilities, agricultural opportunity cost and social vulnerability. Each criterion was standardized to a 0–1 preference scale. Because social vulnerability was a justice concern rather than a simple market benefit, the team reported two MCDA versions: one prioritizing minimum infrastructure cost and another prioritizing equitable access.

Moran's I was calculated for residual flood-risk errors after an initial interpolation. A queen-contiguity matrix produced positive autocorrelation with a permutation p-value below 0.05, indicating that errors were spatially clustered under that specified null model. A k-nearest-neighbour matrix gave a similar conclusion, while a very small distance band created isolated units and was rejected. The team therefore used a spatially informed error model and documented the weights sensitivity.

Three scenarios were evaluated: trend continuation, compact transit-oriented growth and a distributed adaptation scenario. The compact scenario placed more dwellings near transit and reduced agricultural conversion, but required bus-capacity investment. The distributed scenario reduced local congestion in the short term but increased infrastructure extension and fragmented habitat. The adaptation scenario retained larger flood buffers and directed growth toward higher ground, increasing initial land-acquisition costs.

Outcome

No single scenario dominated all criteria. The authority adopted a staged strategy: protect floodways immediately, concentrate early development around two transit nodes, reserve a higher-ground expansion area, and require capacity triggers before later phases. The decision package included maps of robust areas selected across weight ranges, a list of excluded parcels with reasons, and monitoring indicators for flood exposure, housing affordability, transit use and agricultural conversion.

Learning points

The case demonstrates that geospatial analysis is a chain of assumptions. Data quality, spatial weights, normalization, stakeholder values and model structure all influence the result. A useful decision is therefore not merely the highest-scoring map; it is a defensible combination of evidence, uncertainty, safeguards and monitoring.

7. Limitations, Ethics and Governance

Uncertainty and scale

The modifiable areal unit problem occurs when results change with the boundaries or scale of spatial units. The ecological fallacy occurs when relationships observed for areas are attributed to individuals. Positional error, classification error, missing data, temporal mismatch and uncertain flood models can all affect conclusions. Uncertainty should be quantified where possible and described where it cannot be quantified.

Privacy and surveillance

Fine-resolution mobility, property and demographic data may identify individuals or vulnerable groups. Data minimization, aggregation, access control, retention limits and privacy-preserving release are essential. Consent and legal authority should be considered before using location traces. A map can cause harm by exposing informal settlements, shelters or protected persons.

Fairness and participation

A model trained on historical investment may reproduce historical disadvantage. Accessibility metrics should be disaggregated by income, disability, age, gender and other relevant groups when lawful and appropriate. Participation should occur during problem definition and criterion selection, not only after a technical result is complete. Community knowledge can identify omissions and challenge apparently neutral assumptions.

Regulation and accountability

Planning decisions must comply with land-use law, environmental assessment requirements, accessibility obligations, records legislation and data-protection rules applicable in the jurisdiction. Automated or AI-assisted recommendations should remain contestable. Agencies should retain an audit trail showing data sources, model versions, assumptions, decision-makers and reasons for departure from model outputs.

8. Contemporary Developments and Best Practice

Cloud geospatial platforms, open data cubes, machine learning, digital twins, volunteered geographic information and three-dimensional city models are expanding analytical capacity. Machine learning is useful for image classification, demand estimation and anomaly detection, but predictive performance does not guarantee causal validity or fairness. Training data, transferability and explainability must be assessed.

Digital twins can connect real-time sensors, infrastructure models and scenario engines, but they are not automatically complete or accurate representations of cities. Their governance should specify update frequency, ownership, cybersecurity, model uncertainty and permitted uses.

Best practice includes:

• begin with a clearly framed planning decision;
• use the simplest defensible model;
• document data lineage and coordinate systems;
• separate hard constraints from preferences;
• justify weights and test sensitivity;
• test spatial-statistical results under alternative weights matrices;
• distinguish statistical association from causation;
• validate classifications and models with independent evidence;
• report uncertainty, excluded alternatives and distributional effects;
• make outputs reproducible and accessible;
• combine technical analysis with deliberation and post-implementation monitoring.

Chapter Summary

Geospatial analytics integrates spatial data, statistics, modelling and institutional judgement to support urban and regional decisions. GIS and remote sensing provide structured evidence, but coordinate systems, metadata, classification accuracy and temporal consistency determine fitness for use. Spatial statistics require an explicit spatial-weights matrix and careful interpretation of significance; a significant Moran's I indicates evidence against a specified randomization hypothesis, not causation.

Suitability analysis combines constraints and standardized criteria. MCDA makes preferences visible through weights and aggregation, but rankings must be tested for sensitivity and double-counting. Scenario planning explores coherent alternative futures rather than claiming certainty. Urban models—including interaction, land-use–transport, cellular automata, agent-based and system-dynamics models—are conditional representations requiring calibration, verification and validation.

Responsible decision support communicates assumptions, uncertainty and distributional effects. Ethical practice protects privacy, challenges historical bias, enables participation and preserves human accountability. The strongest planning recommendations are reproducible, transparent, robust across plausible assumptions and linked to monitoring after implementation.

Key Terms

Spatial database: A database designed to store, query and manage geometry together with attributes.

Remote sensing: Observation of Earth's surface from sensors located away from the target.

Raster: Spatial data represented as a regular grid of cells.

Vector: Spatial data represented by points, lines and polygons.

Spatial autocorrelation: Dependence among values associated with nearby locations.

Spatial weights matrix: A matrix expressing the assumed neighbourhood or influence relationship among locations.

Row standardization: Rescaling weights so each non-zero row sums to one.

Moran's I: A global statistic for assessing spatial autocorrelation.

Permutation test: A significance procedure based on repeatedly rearranging observed values under a null hypothesis.

Spatial heterogeneity: Variation in relationships or processes across geographic space.

Suitability analysis: Evaluation of locations against constraints and preference criteria.

Constraint: A condition that excludes an alternative or location from consideration.

Standardization: Transformation of measurements to a common comparison scale.

MCDA: Multicriteria decision analysis using multiple criteria to compare alternatives.

Sensitivity analysis: Examination of how results change when assumptions or inputs vary.

Scenario planning: Exploration of alternative coherent futures under uncertainty.

Calibration: Estimation of model parameters using observations.

Validation: Assessment of model performance against independent or relevant evidence.

Spatial lag: A weighted combination of values at neighbouring locations.

Digital twin: A data-linked digital representation of a physical system used for monitoring or simulation.

Review Questions

Short Answer Questions

1. Distinguish raster and vector data and give one planning application for each.

2. What does row standardization do to a spatial weights matrix?

3. Why is a significant Moran's I not evidence of causation?

4. Distinguish a constraint from a criterion in suitability analysis.

5. Explain why scenarios should not be described as forecasts.

Long Answer Questions

1. Explain the construction, interpretation and limitations of a spatial weights matrix and Moran's I.

2. Describe a complete MCDA workflow for selecting a new public facility site.

3. Compare land-use–transport, cellular-automata and agent-based urban models.

4. Discuss uncertainty, privacy and fairness in geospatial decision support.

5. Explain how calibration, verification and validation differ in urban modelling.

Analytical / Application Questions

1. Using the worked MCDA data, calculate the score of each site and test how the ranking changes when the access weight is increased to 0.50 and the environmental weight is reduced to 0.25.

2. Design a permutation-based Moran's I analysis for neighbourhood-level vacancy rates. Specify the weights matrix, null hypothesis, test procedure and interpretation of a p-value of 0.03.

3. Develop three internally coherent scenarios for a rapidly growing coastal city facing flood risk. Identify criteria, constraints, indicators and a monitoring plan for comparing them.

Multiple Choice Questions with Answers

Question 1

Which statement best describes a row-standardized spatial weights matrix?

A. Every diagonal element is set to one

B. Each non-zero row is rescaled to sum to one

C. All pairwise distances are converted to kilometres

D. Every location is assigned the same number of neighbours

Correct Answer: B. Each non-zero row is rescaled to sum to one

Question 2

For the stated randomization null hypothesis, a significantly positive Moran's I most directly indicates:

A. Clustering of similar values under the specified spatial weights

B. A causal effect of the measured variable on its neighbours

C. That all observations are statistically independent

D. That the regression model has no omitted variables

Correct Answer: A. Clustering of similar values under the specified spatial weights

Question 3

Using the worked example, which site has the highest original MCDA score?

A. Site A, with 0.705

B. Site B, with 0.7175

C. Site C, with 0.635

D. All sites tie at 0.700

Correct Answer: B. Site B, with 0.7175

Question 4

In the worked MCDA example, what is the approximate margin between Site B and Site A?

A. 0.0015

B. 0.0125

C. 0.0500

D. 0.0825

Correct Answer: B. 0.0125

Question 5

Which operation is most appropriate for a legally protected wetland in a suitability model?

A. Treat it as a hard exclusion constraint

B. Assign it the average preference score

C. Duplicate it as two positive criteria

D. Ignore it until final map design

Correct Answer: A. Treat it as a hard exclusion constraint

Question 6

A permutation test for Moran's I primarily generates a reference distribution by:

A. Changing the map projection repeatedly

B. Randomly rearranging observed values among fixed locations

C. Deleting all outliers from the dataset

D. Increasing the number of spatial units

Correct Answer: B. Randomly rearranging observed values among fixed locations

Question 7

If a two-sided permutation p-value is 0.03 and alpha is 0.05, the appropriate conclusion is:

A. Reject the specified spatial-randomness null hypothesis

B. Prove that the observed pattern is caused by the weights matrix

C. Accept that no spatial pattern exists

D. Prove that every local area is significant

Correct Answer: A. Reject the specified spatial-randomness null hypothesis

Question 8

Which statement correctly distinguishes calibration from validation?

A. Calibration estimates parameters; validation assesses performance against relevant evidence

B. Calibration checks computer code; validation selects a coordinate system

C. Calibration removes spatial dependence; validation creates survey data

D. Calibration is performed only after implementation; validation is never empirical

Correct Answer: A. Calibration estimates parameters; validation assesses performance against relevant evidence

Question 9

What is a principal limitation of an additive MCDA model?

A. It cannot use more than one criterion

B. It assumes criteria are sufficiently compensatory

C. It always produces statistically significant results

D. It requires all inputs to be measured in metres

Correct Answer: B. It assumes criteria are sufficiently compensatory

Question 10

Which statement about scenario planning is most accurate?

A. It identifies one certain future from historical trends

B. It explores coherent alternative futures and tests policy robustness

C. It eliminates the need for stakeholder participation

D. It is equivalent to classifying satellite imagery

Correct Answer: B. It explores coherent alternative futures and tests policy robustness