eduKateSG · Hydrology · The Gold Standard of Education
Turn yesterday’s rain into a testable memory, not a hidden certainty
Choose the explanation you need, or follow all twelve chapters in order.
Full chapter index · Independent practice · Hydrology subject guide
Two storms can deliver the same rainfall depth and produce different runoff. The catchment may begin one storm dry and the other wet; storage may be empty or full; vegetation, season, groundwater and drainage may differ. An antecedent precipitation index, or API, compresses recent rainfall history into a decaying numerical state.
The index is useful because it makes one kind of memory explicit. It is limited because rainfall history is not the same thing as measured soil moisture, groundwater state or total catchment storage. The gold standard treats API as a model with a defined update rule, calibration period and intended response variable.
These chapters build an API day by day, connect it to runoff response, and test what changes when the time step, decay factor, rainfall record or catchment changes. Worked values are original teaching examples. Official hydrology sources support the specialist principles; they do not validate the invented coefficients or fictional flow records.
Continue from The Gold Standard of Hydrology within The Gold Standard of Education: learning that transfers into life and the How X Works learning library.
Full chapter index
See the mechanism · Chapters 1–2
Build the worked model · Chapters 3–5
Test what changes · Chapters 6–8
Check evidence and limits · Chapters 9–10
Practise and continue · Chapters 11–12
Begin with the hydrologic question. Are you trying to estimate the initial infiltration condition for an event, explain why two similar storms produced different peaks, initialise a model, or screen days for likely wetness? A useful memory variable must be matched to the process and time scale.
An API is commonly written as a recursive balance such as API today = rainfall today + k × API yesterday. The decay factor k, between zero and one in the simple form, controls how quickly earlier rainfall loses influence. A larger k creates a longer mathematical memory.
This recursion is not a direct measurement of water stored everywhere in the catchment. It is a weighted history of recorded precipitation. The relation to physical wetness must be established empirically or through an appropriate model.
State the time step. A daily index and an hourly index cannot share the same numerical k without changing the implied decay time. State whether rainfall is added before or after decay; both conventions exist, and silent switching changes the series.
State the precipitation input too. One gauge, a weighted catchment mean, radar rainfall and a merged product can give different histories. A precise formula applied to unrepresentative rainfall remains a poor catchment memory.
The USGS rainfall–runoff simulation model report includes an antecedent-precipitation index used in estimating an initial infiltration condition. Treat that as a documented application, not a universal coefficient for every basin.
CHAPTER 2 OF 12 · See the mechanism
2. Read exponential decay as a sequence of weights
Suppose k = 0.80 and no new rain falls. Yesterday’s API contributes 80% today, 64% after two days, 51.2% after three and 40.96% after four. Each step multiplies the previous contribution by the same factor.
Expand the recursion. Today’s API equals today’s rain plus 0.80 times yesterday’s rain, plus 0.80² times rain two days ago, and so on. This reveals the index as an exponentially weighted sum rather than a mysterious state variable.
Compare k = 0.50 and k = 0.95. With 0.50, rainfall influence halves each step. With 0.95, it fades slowly. Neither is “more accurate” without a target process, data and calibration. The factor encodes a memory assumption.
A half-life can make the assumption easier to discuss. Solve k^h = 0.5 for h steps. For k = 0.80, h is about 3.1 days. This is a mathematical half-life of the rainfall contribution, not necessarily the drainage half-life of a soil or aquifer.
Seasonal evapotranspiration, snow, irrigation or controlled drainage may violate a constant decay assumption. A single k can still be a useful summary, but its limits should be visible beside the result.
Ask learners to draw the weights of rainfall from each of the last ten days. The shape makes clear why a large storm can remain influential and why a small recent shower can still alter the current index.
CHAPTER 3 OF 12 · Build the worked model
3. Initialise the index without hiding the starting condition
A recursive index needs a starting value. Setting API to zero on the first day is convenient, but it assumes no relevant earlier rainfall memory. If the record begins during a wet spell, early values can be biased low.
One remedy is a warm-up period. Begin the calculation earlier than the analysis window and discard the initial portion. Another is to estimate the initial state from preceding rainfall or an observed variable. The appropriate choice depends on data availability and sensitivity.
Document missing rainfall. Treating a missing day as zero says “no rain,” which is a hydrologic claim. If the gauge failed, the correct state is unknown until repaired or represented with an uncertainty rule.
Gauge changes can create artificial jumps. A station may move, a sensor may be replaced, or reporting intervals may change. Quality control belongs before the recursion because every input affects later days through the memory term.
For a catchment mean, use a documented weighting method. The existing Thiessen Rainfall Weights and Gauge Area guide explains one geometric method and its limits. A fixed polygon weight does not correct storm-specific spatial structure.
Perform an initial-condition sensitivity check. Recalculate with plausible starting values. If conclusions converge after several decay times, the warm-up may be adequate. If not, report that the result depends materially on an unobserved starting state.
Use API_t = P_t + 0.80 API_{t-1}, with API before day one set to zero. Day one rain is 18 mm, so API becomes 18.00. Day two is dry: API becomes 0 + 0.80 × 18 = 14.40.
Day three brings 6 mm. API becomes 6 + 0.80 × 14.40 = 17.52. Day four is dry, giving 0.80 × 17.52 = 14.016, rounded to 14.02 mm-equivalent units in the table.
The unit resembles rainfall depth because the index is a weighted sum of depth. It should still be labelled an index. API = 14.02 does not assert that 14.02 mm of water is stored in the soil profile.
Change k to 0.60. Day two becomes 10.80, day three 12.48 and day four 7.49. The lower k forgets the day-one storm more quickly. The arithmetic difference expresses a model choice, not new weather.
Now shift the time step to hours. A coefficient of 0.80 per hour would forget far more rapidly over a day than 0.80 per day. Convert decay consistently or recalibrate against hourly responses.
Check the sequence by expanding day four: 0.80³ × 18 + 0.80 × 6 = 14.016. This second route catches update-order mistakes and shows the separate contribution of each rainfall event.
| Day | Rainfall (mm) | Previous API | New API |
|---|---|---|---|
| 1 | 18.0 | 0.0 | 18.00 |
| 2 | 0.0 | 18.0 | 14.40 |
| 3 | 6.0 | 14.4 | 17.52 |
| 4 | 0.0 | 17.52 | 14.02 |
CHAPTER 5 OF 12 · Build the worked model
5. Connect API to runoff without claiming causation too quickly
A high API may be associated with larger runoff because wetter antecedent conditions can reduce available storage or change connectivity. The direction is physically plausible, but the relationship must be tested with event data and an appropriate response measure.
Choose the response: peak discharge, runoff volume, runoff coefficient, time to peak or threshold exceedance. Different parts of the hydrograph can relate differently to antecedent state. A single correlation with “runoff” can hide this distinction.
Control the event definition. If one analysis separates storms by six dry hours and another by twenty-four, rainfall totals and antecedent windows change. State how events begin and end.
Plot the response against both event rainfall and pre-event API. A large storm on a dry catchment can exceed a small storm on a wet catchment. The index is one explanatory dimension, not a replacement for event magnitude and intensity.
Look for interactions. Antecedent wetness may matter more for moderate storms than for extreme storms that overwhelm storage. Alternatively, a high-intensity burst may generate runoff on a dry surface through a different mechanism.
The causal story must respect measurement. If API and event rainfall share the same gauge errors, apparent relationships can be distorted. Use independent soil-moisture or flow observations where feasible to test the proposed meaning.
CHAPTER 6 OF 12 · Test what changes
6. Distinguish API from soil moisture, groundwater and storage
Soil moisture is a physical state measured or modelled for a specified depth, location and method. Groundwater level is another state with its own response time. Surface depression storage, wetlands and reservoirs add further memory. API summarises rainfall history and cannot silently stand for all of them.
The USGS-hosted study of soil-moisture initialisation discusses antecedent soil moisture as catchment memory in model initialisation. That supports the distinction: measured or estimated moisture provides different information from rainfall-only memory.
Two periods can share the same API but differ in temperature, vegetation, drainage or groundwater. If those differences matter, the same index value can produce different runoff responses. This is model non-uniqueness, not necessarily calculation error.
API can still be valuable when direct state observations are sparse. Its inputs may be widely available and its update is transparent. The correct claim is that it is a proxy whose usefulness has been evaluated for a particular purpose.
Compare API with soil-moisture observations over time. Examine lag, season and dry-down. A good association in one season may weaken in another when evapotranspiration or frozen conditions alter the water balance.
When combining state variables in a model, avoid double-counting the same memory. If soil moisture already carries rainfall effects, adding API as an independent predictor may inflate apparent information unless the model and validation support it.
CHAPTER 7 OF 12 · Test what changes
7. Compare two storms with the same event rainfall
The fictional table gives both storms 42 mm of event rainfall. Storm A begins with API 8 and produces an 18 m³/s peak and 52,000 m³ runoff volume. Storm B begins with API 31 and produces a 39 m³/s peak and 111,000 m³ volume.
The comparison is consistent with a wetter antecedent state increasing response, but two observations cannot establish a general relationship. Intensity pattern, storm movement, season and measurement error could also differ.
Calculate runoff ratios only if the catchment area and rainfall representation are suitable. A volume cannot be compared with depth without converting units. A single point gauge may not represent the rainfall over the contributing area.
Add matched storms across many API levels. Use the same event-separation rule and quality checks. Plot residuals after accounting for event depth or intensity. If API explains a consistent part of the remaining variation, the memory measure gains support.
Test out-of-sample performance. A coefficient chosen on all events can appear impressive because it was tuned to them. Reserve a later period or use cross-validation that respects time order.
The changed-context question is decisive: would the interpretation survive if Storm B had a lower peak but larger volume? That result could reflect slower pathways or storage release. “Wetter means bigger” is too coarse to explain hydrograph shape.
| Storm | Event rain | API before storm | Peak flow | Runoff volume |
|---|---|---|---|---|
| A | 42 mm | 8 | 18 m³/s | 52,000 m³ |
| B | 42 mm | 31 | 39 m³/s | 111,000 m³ |
CHAPTER 8 OF 12 · Test what changes
8. Change the decay factor, window and spatial scale
Fit k by minimising an error against a target response, or select it from a physically justified timescale. In both cases, report the method. Trying many coefficients and presenting only the best one without validation exaggerates certainty.
Test a family of k values. If model performance is nearly flat from 0.75 to 0.90, the exact optimum is weakly identified. Report the stable range rather than treating the third decimal place as a catchment constant.
A finite moving-window index is another option: sum rainfall over the previous seven days, perhaps with equal or custom weights. It has a sharp boundary where day eight contributes nothing. Exponential decay has no abrupt cut-off but relies on its functional form.
Spatial scale changes the meaning. A field-scale API may relate to a local soil response; a basin-scale index may need distributed rainfall and subcatchment states. Averaging first and then applying a nonlinear runoff model can differ from modelling subareas first.
Land-use change can break a historical calibration. New impervious surfaces, drainage works or vegetation changes alter how a given rainfall memory translates into runoff. The index series may remain arithmetically identical while its predictive relationship changes.
Review k after known changes and through rolling validation. Stability is evidence; forced constancy is not. A model should be allowed to reveal when yesterday no longer predicts today in the same way.
CHAPTER 9 OF 12 · Check evidence and limits
9. Audit rainfall, flow and event records before calibration
Inspect rainfall timestamps, accumulation intervals and time zones. A daily gauge ending at 09:00 does not align automatically with a midnight-to-midnight flow day. Misalignment can make antecedent rain look like event rain or vice versa.
Check gauge consistency and missing periods. The existing Double-Mass Curves and Rain-Gauge Consistency guide provides a route for detecting shifts relative to neighbouring records, while recognising that regional climate variation can also change slopes.
Inspect discharge rating changes and flagged values. A flow series derived from stage inherits uncertainty from the rating curve, especially outside measured ranges. Do not calibrate a precise wetness threshold against unexamined flow estimates.
Separate snow, tropical convective and frontal-rain contexts when their processes differ. An index designed for liquid rainfall and warm soils may fail where snow accumulation and melt control the timing.
Record human operations. Reservoir releases, pumps and gates can produce hydrograph changes unrelated to antecedent precipitation. Naturalised or operationally annotated flows may be necessary for the stated question.
The audit result should be part of the model record: usable periods, repairs, exclusions and remaining uncertainties. Clean-looking graphs are not a substitute for traceable inputs.
CHAPTER 10 OF 12 · Check evidence and limits
10. State limits and avoid common API misconceptions
First misconception: API is measured soil moisture. It is not. It is a rainfall-memory index whose relationship with moisture must be established for the relevant location, depth and season.
Second misconception: one decay factor is a universal catchment property. The fitted value depends on time step, input data, target variable, period and model form. It may also change through the year.
Third misconception: high API guarantees flooding. Flood response depends on event rainfall, intensity, spatial pattern, drainage capacity, storage, operations and many other conditions. API can modify risk without determining the outcome alone.
Fourth misconception: dry days contain no information. They drive decay and therefore update the assumed memory. Missing dry-day records must not be silently treated as verified zero rainfall.
Fifth misconception: better calibration proves physical truth. A flexible relationship may predict well without representing the actual storage mechanism. Prediction and explanation are related but distinct goals.
Finally, an API does not replace current warnings, operational hydrology or professional flood assessment. It is an educational and modelling tool. Decisions affecting safety require current observations, responsible authorities and fit-for-purpose systems.
CHAPTER 11 OF 12 · Practise and continue
11. Practise with a transparent API notebook
Create a ten-day rainfall series containing one large storm, two small showers and several dry days. Calculate API for k = 0.60, 0.80 and 0.95. Plot the three curves and explain which past event dominates each day.
Change the initial API from zero to 20. Mark when the difference between the two runs becomes negligible for the decision. This demonstrates warm-up rather than merely defining it.
Delete one rainfall value. Compare three treatments: zero, interpolation and unknown. Explain why zero is defensible only when evidence supports no rain. Preserve the original missing flag even if a repaired series is produced.
Pair the API with fictional peak-flow values. Fit no more than a simple relationship, then reserve the last events for checking. Report whether the coefficient improves predictions beyond event rainfall alone.
Questions learners often ask Can k exceed one? Not in the simple decaying-memory form used here; that would amplify old rainfall. Is API a probability? No. Does a higher API always mean a higher peak? No. Can radar rainfall be used? Yes, if its processing, representativeness and errors are appropriate and documented.
Finish with a model card: formula, update order, time step, rainfall source, initialisation, calibration target, valid period and review trigger. If another reader cannot reproduce the index, the memory has not been made operational.
CHAPTER 12 OF 12 · Practise and continue
12. Connect rainfall memory to the wider Hydrology route
Return to The Gold Standard of Hydrology for precipitation, infiltration, storage, flow, measurement and model limits. API is one compact route through that larger system.
Use Lag Time, Time of Concentration and Catchment Clocks when the question concerns travel and hydrograph timing. Antecedent memory and response timing interact, but they are not the same clock.
Use Unit Hydrographs, Shift-and-Add Limits when modelling event response under linear assumptions. API may help define initial conditions; it does not make a changing catchment linear.
The USGS sources linked above were checked on 22 October 2026. They support documented uses of antecedent precipitation and soil-moisture initialisation. The daily formula, coefficients, storm flows and tables here are original examples.
The main limit is purposeful: this article does not supply a universal flood threshold or operational forecast. A real application needs local calibration, independent validation, quality-controlled rainfall and flow, and review after environmental or infrastructure change.
The durable transfer is to expose memory as an assumption. Write how yesterday enters today, test how quickly it fades, and check whether the chosen memory predicts the response outside the data that created it.
A final calibration and validation route Split the workflow into calibration, validation and monitoring. Calibration chooses k and any regression coefficients. Validation applies those fixed choices to a separate period. Monitoring checks whether errors, input quality or catchment conditions later move beyond agreed limits. Re-estimating coefficients every time performance falls is not an independent test.
Use time-aware validation. Randomly scattering neighbouring days between training and test sets can leak the same wet spell into both. A blocked later period, several withheld seasons or leave-one-event-out design may better reflect the intended prediction task.
Choose performance measures that match the decision. Mean squared error emphasises large misses; absolute error is less dominated by extremes; event-detection scores examine thresholds; bias shows systematic over- or underprediction. One summary number cannot show every failure mode.
Plot errors against API, event depth, season and time. If the model underpredicts consistently when API is high, the relationship may be nonlinear or storage may saturate. If errors drift after drainage works, a structural change is more plausible than random noise.
Compare the API model with simple baselines. A predictor using only event rainfall, seasonal average or yesterday’s flow may perform surprisingly well. The index earns its place when it adds useful out-of-sample information, not merely when its fitted coefficient has the expected sign.
Quantify observational uncertainty where possible. Rain-gauge undercatch, spatial sampling, rating-curve error and missing data can widen the plausible response. A narrow model interval based only on regression residuals may ignore the largest measurement uncertainties.
For operational use, define failure behaviour. If rainfall input stops, should the system freeze the last API, continue decay, mark the state unavailable or switch to another data source? Each choice makes a different claim. The safe rule should be documented before an outage.
Archive every model version with its data window, formula, parameters and verification result. A later user should be able to reconstruct why an alert or analysis used a particular state. Without versioning, the same date can acquire different API values after a silent recalibration.
Treat surprising storms as learning opportunities, not automatic reasons to discard data. Inspect whether the event reveals a new process, a measurement failure or a legitimate tail case. Update only after preserving the original forecast and the evidence that challenged it.
Check parameter transfer between catchments. Similar area does not imply similar soils, relief, drainage density, storage or rainfall organisation. If a nearby calibrated basin supplies k, label the regionalisation rule and test it against whatever local observations exist.
Consider multiple memories when the process supports them. A fast index can represent near-surface response while a slow index tracks deeper or seasonal storage. Extra states increase flexibility and the risk of weak identification, so validate whether both improve prediction for the stated task.
Communicate uncertainty in ordinary language. Instead of presenting API 27 as a measured wetness, say that recent rainfall history is high under this model and that the relationship to runoff was calibrated for a named period. Readers can then separate the computation from the decision built upon it.
Keep the raw precipitation series beside the derived index. Derived states can be recalculated when an input is corrected, a coefficient changes or a new spatial weighting becomes available. Storing only API values makes quality repair difficult and can leave later analysts unable to distinguish a wet day from a long memory of earlier rainfall.
Document threshold use separately from index construction. A threshold chosen to flag high-response events depends on loss tolerance, sample prevalence and operational lead time. The same API formula can support different thresholds for research, screening and warning, but those purposes should never be blended silently.
Recheck the index after corrected rainfall arrives. Preserve both the value available in real time and the later quality-controlled value, because forecast evaluation and historical analysis answer different questions. This distinction prevents hindsight data from making an operational model look better than it was when the decision had to be made.