Data & Model Evaluation (agsy-data-and-model-evaluation)

A model is only as credible as its evaluation. AgSy reviewers are systems-modelling experts: they want to see the model tested against independent data, its sensitivity and uncertainty characterized, and the trade-offs the system exhibits — not a single tuned run presented as truth. Model description and choice live in agsy-systems-framing-and-modeling; this skill covers testing and reporting.

When to trigger

• Reporting how well the model reproduces observations
• Running sensitivity / uncertainty analysis
• Building scenario comparisons and trade-off analyses
• A reviewer asked for validation, sensitivity, uncertainty, or alternative scenarios

Evaluation norms AgSy expects

1. Independent evaluation. Compare observed vs. simulated on data not used for calibration. Report standard fit statistics — RMSE, RRMSE, bias/ME, modelling efficiency (NSE), index of agreement (d), R² — and show the 1:1 plot. State what "good enough" means for the decision.
2. Sensitivity analysis. Identify the parameters/inputs that drive outputs (local one-at-a-time and, where feasible, global methods — Morris, Sobol). Report which assumptions matter most.
3. Uncertainty. Propagate uncertainty from inputs, parameters, and structure to the outputs that carry the conclusions. Present ranges/intervals, not point estimates dressed as certainty.
4. Trade-off & scenario analysis. This is where AgSy papers earn their place: compare scenarios or management/design options on multiple objectives (yield, profit, environment, risk) and show the trade-offs and synergies — Pareto fronts, trade-off curves, multi-indicator profiles.
5. Scaling. When you aggregate from field to farm to region, state how, and check that aggregation does not hide compensating errors.

Stochastic & data-driven components

• For Monte Carlo, ABM, or stochastic weather/price generators: set and report seeds, run enough replicates, and report the distribution, not one realization.
• For surrogate/emulator or ML components: validate against the full model and report where they fail.

Reproducibility while you work (not at the end)

• One master workflow regenerates every figure/table from inputs + model runs.
• Pin model version, parameter sets, and software versions; record calibration vs. evaluation splits.
• Keep table/figure numbers matched to script/model outputs — the data/code/model are deposited (see agsy-reproducibility-and-data-policy).

Anti-patterns

• Calibrating and "validating" on the same data; reporting fit only on the training period
• A single tuned run with no sensitivity or uncertainty analysis
• Reporting R² alone (insensitive to bias) — pair it with RMSE/NSE/bias
• Scenario tables with no trade-offs surfaced (just "scenario B is best")
• Aggregating across scales without checking for compensating errors

Evaluation completeness rubric (what an AgSy referee checks)

A systems referee scores the model section on five axes (strong → weak, a weak answer often draws a major revision): independence (held-out sites/years → "validated" on calibration data); fit (RMSE+NSE+ bias+1:1 plot → R² only); sensitivity (global Morris/Sobol → one-at-a-time); uncertainty (propagated intervals → a discussion mention); trade-offs (Pareto front → a "best scenario" sentence).

Worked micro-example (illustrative numbers)

A bioeconomic whole-farm model evaluates climate-adaptation options for a mixed crop–livestock farm. Numbers are illustrative — not real results.

• Evaluation. Calibrated on 2008–2015; held-out 2016–2020 gives margin RMSE = €185/ha, NSE = 0.74, bias = −€20/ha with the 1:1 plot; "good enough" is RRMSE < 15%, and the negative bias is flagged.
• Sensitivity. Sobol indices: rainfall (0.41) and price (0.27) dominate, so results are reported on a rainfall/price grid, not a single forecast.
• Trade-off. Option B raises margin +8% but N surplus +14%; D loses 3% margin but cuts N surplus 22%. The Pareto front is plotted; the recommendation names the trade-off, not a single winner.

Referee pushback → the AgSy-specific fix

• "Not validated against independent data." → Hold out sites/years, report the full fit panel on the held-out set, and state the decision-relevant tolerance.
• "Trade-offs across objectives not quantified." → Replace "best scenario" with a multi-objective table or Pareto front showing what each option gains and gives up.

Calibration anchors (hedged where policy is volatile)

• Fit-statistic conventions (RMSE, RRMSE, NSE, bias, d) are community norms, not a fixed journal rule.
• "Independent evaluation" means data withheld from calibration; the exact bar is judged by referees.
• Whether a data/code/model deposit is required versus encouraged is an Elsevier research-data question — confirm against the journal's current author guidelines.

Output format

【Evaluation data】independent of calibration? [Y/N]
【Fit statistics】RMSE / NSE / bias / d / R² + 1:1 plot
【Sensitivity】which inputs/parameters dominate
【Uncertainty】propagated to the conclusion-bearing outputs? [Y/N]
【Trade-offs】scenarios compared on multiple objectives (Pareto/curve)
【Reproducible】master workflow + pinned model/version + seeds? [Y/N]
【Next】agsy-figures-and-tables

Supplementary resources

• ../../resources/external_tools.md — calibration, sensitivity, uncertainty, and multi-objective packages
• ../../resources/official-source-map.md — research-data/model reproducibility policy