
test forecast methods against the past, judge how data limits and bias affect my interpretation, and use critique to improve the forecast card.
Phase 4: Build
Do Now. Pretend it is the year 2000. Using only the 1959 and 2000 Mauna Loa values, forecast 2024. Write your number before you look at the real one.
The 1959–2000 rate is about 1.31 ppm per year. What does a straight line from 2000 predict for 2024?
Hindcast test. A method that cannot reproduce the past should not be trusted with the future. Hide the recent data, forecast it, then compare.
| Method (trained on data up to) | Predicted 2024 (ppm) | Observed 2024 (ppm) | Error |
|---|---|---|---|
| A: long line, 1959–2000 rate (1.31 per year) | 401.2 | 424.61 | −23.4 (−5.5%) |
| B: recent rate, 2000–2014 (2.08 per year) | 419.6 | 424.61 | −5.0 |
| C: class model from 2014, AF 0.45, 11 GtC a year | 422.0 | 424.61 | −2.6 |
| C revised: AF 0.50 | 424.6 | 424.61 | about 0 |
Methods A and B use the real Mauna Loa values. Method C assumes 11 GtC of emissions every year, a simplification; real emissions rose slowly over the decade.
Worked example 1: reading the test. Every method was too low, and the simplest was worst. That is the signature of an accelerating quantity: the longer the training period, the slower the average rate and the bigger the miss. The revised model fits best, but it was tuned to this very decade, so its good fit is weaker evidence than a fit to data it never saw.
Worked example 2: data limits and bias. The harbor forecast rests on one tide gauge. If the pier under it has been sinking 0.5 mm a year, the gauge overstates the ocean’s rise by that much, but that sinking is real for the city too, so for planning it still counts. If the gauge was moved in 1950 and the record was not corrected, rates across 1950 are biased. A careful forecaster checks a second gauge or satellite data and says what was checked.
Gallery critique. Cards go on the wall. Visitors leave notes in three colors: I like (what works, with evidence), I wonder (a question about a range, method or source), What if (a specific change). Then each team tests V1 against the criteria checklist and makes version 2.
Reversible or not? Version 2 must say which changes could be reversed by 2100 if emissions fell sharply (the CO2 growth rate, part of the warming trend) and which could not (most of the sea-level rise already under way, and the ocean’s lower pH for a long time).
Run the gallery, then test and revise.
Make version 2 and justify it.
Make Forecast Card V2. Under it, list each change, the critique note or failed criterion that caused it, and whether the change made the card more accurate, more honest about uncertainty, or easier to read.
Write the card's “limits” box: one data limitation or possible bias, one model limitation, how each could shift your 2050 range, and which predicted change is reversible and which is not.
London is protected from North Sea storm surges by the Thames Barrier, a row of movable gates across the River Thames that became operational in 1982. It was built after a deadly 1953 North Sea flood, and it closes when a high tide and a storm surge are forecast together. Because the barrier and the river’s walls were not designed for the higher seas expected later this century, planners wrote a long-term plan for the estuary to the year 2100. Instead of betting on one sea-level number, the plan uses a range of projections, watches a set of indicators such as measured sea level and how often the barrier closes, and sets decision points: when the indicators reach agreed levels, the next upgrade is started. The plan is reviewed as new evidence arrives.
Think about it: How does planning around a range with decision points help a harbor city act now without knowing exactly how fast the sea will rise?
Version 2 has faced the past and the gallery. Next phase: the full written forecast and its defense.
How confident are you that you can I can test a forecast method against the past and explain how data limits affect my conclusions.
Exit ticket: which critique note changed your card most, and what did the change do to your range?
Materials: V1 cards, sticky notes in three colors, the hindcast table, criteria checklists, calculators.
Setup: run the gallery for 12 minutes, then testing and V2. Collect V1 and V2 for the 3.09 forecast.
Facts to keep careful: the case study names only the barrier’s start of operation (1982), the 1953 flood and the 2100 planning horizon; check current sources before adding numbers.
Listen for: V2 changes with no link to a critique; “the fit proves it”; bias confused with random error.
Support / stretch: give a two-option menu of V2 changes. Stretch: repeat the hindcast for 2014 using 1959–1990 data from the full NOAA record.