1. Time order is information
A time series records a variable in chronological order at defined intervals. Daily withdrawals, monthly deposits and quarterly operating expenses are examples. The ordering matters because neighbouring observations may be related and calendar patterns may repeat. Rearranging the observations destroys information about trend and timing. A cross-sectional comparison of many branches on one date answers a different question from a series for one branch across many dates. A panel combines both dimensions, but this chapter focuses on a single ordered series.
Begin by defining what each observation represents. A month-end deposit balance is a stock measured at a point in time. Monthly transaction volume is a flow accumulated over an interval. Summing month-end balances does not produce annual deposits in the same sense as summing monthly transaction counts produces annual transactions. Record the frequency, units, coverage and treatment of missing periods. A missing month is not a zero-volume month. Changes in branch coverage or definitions can create apparent growth that is unrelated to underlying business performance. Plot the original series before choosing a forecasting method so unusual breaks, seasonality and data issues become visible.
2. Trend, seasonal, cyclical and irregular movement
Trend is a persistent long-run movement in the level of a series. Seasonality repeats at a known calendar frequency, such as higher payment volume around salary dates or specific months. Cyclical movement refers to longer fluctuations whose duration need not be fixed. Irregular movement includes unusual shocks and residual variation not explained by the other components. These categories help organise thinking, but the separation is not uniquely observable without modelling choices and sufficient history.
A payment series can show a rising trend as digital adoption expands, a monthly seasonal pattern around payroll dates and an irregular drop during an outage. A longer business cycle may influence credit demand over several years. Do not call every rise a trend or every oscillation seasonal. A single year of monthly observations may reveal within-year differences but cannot establish stable recurrence across years. The observation frequency also matters. Daily data can show weekday effects that disappear in monthly totals. Identifying components helps choose a method, but the model should reflect the decision horizon. Tomorrow's cash requirement and next year's branch capacity may require different summaries of the same underlying activity.
3. Additive and multiplicative decomposition
An additive decomposition writes Y = T + S + C + I. It is useful when seasonal changes have roughly constant absolute size as the series level changes. A multiplicative decomposition writes Y = T × S × C × I and is useful when seasonal movements are proportional to the level. In simpler models, trend and cycle may be combined. The components depend on the decomposition method and are not independent facts read directly from the observed series.
Suppose a series with baseline one hundred increases by twenty in a seasonal peak, and a later baseline two hundred increases by about twenty again. An additive effect is plausible. If the seasonal increase becomes forty at the higher baseline, a multiplicative factor near 1.20 may fit better. Under a multiplicative model, observed volume 120 divided by seasonal factor 1.20 gives a deseasonalised value of one hundred. Under an additive model with seasonal effect twenty, subtract twenty. Do not subtract a multiplicative index or divide by an additive effect. Multiplicative models also need care with zero or negative values, where proportional interpretations or logarithmic transformations may fail.
4. Plotting and preparing the series
Use a line chart with time on the horizontal axis and the measured variable on the vertical axis. Preserve unequal gaps when observations are irregularly spaced, or explicitly aggregate to a regular frequency. Mark changes in definitions, branch additions and exceptional events. A sudden level jump may reflect a system migration or new reporting scope rather than growth. Before smoothing, reconcile totals and verify whether revisions have changed earlier values. A clean time index is as important as clean numerical values.
Calendar adjustment can matter. A month with more working days may naturally have more transactions even when daily activity is unchanged. Compare both totals and activity per working day when the decision requires it. However, do not automatically divide every series by working days. Month-end balances are stocks, and their interpretation differs from transaction flows. If an outage creates a missing value, filling it by interpolation makes an assumption about the unobserved activity. Record that assumption and assess its effect. Historical data used for a realistic forecast evaluation should reflect what was available at the forecast date, including any reporting delays, rather than silently using later revised information.
5. Moving averages for smoothing
A moving average replaces a local set of observations with their mean to reduce short-term variation. A centred three-period average for values ninety, one hundred and one hundred ten is one hundred, assigned to the middle period. As the window moves forward, the earliest value leaves and the next value enters. A longer window smooths more strongly but can conceal turning points and leaves fewer centred values near the ends. The window should reflect the frequency and purpose of the analysis.
Centred smoothing uses observations on both sides of the labelled period. It is useful for describing historical patterns but cannot be used as though all its values were available in real time. A trailing three-period forecast uses the latest three known observations to predict the next period. The arithmetic may match a centred average, but the timing interpretation differs. For known values ninety, one hundred and one hundred ten through period three, their average one hundred can be a simple forecast for period four. It should not be labelled a forecast made before period three was observed. This distinction prevents look-ahead bias in evaluation.
6. Even-period centring
An even-length moving average falls between time points. For quarterly data, a four-quarter average is centred halfway between the second and third quarters in its window. To align it with an actual quarter, average two adjacent four-quarter averages. This additional step produces a centred moving average. It is commonly used when estimating a trend-cycle component before measuring seasonal ratios. Losing some observations at both ends is a consequence of using surrounding data.
Consider five consecutive values eighty, one hundred, one hundred twenty, one hundred and one hundred. The first four-period average is one hundred. The next is 105. Their average is 102.5 and is aligned with the third observation. If that observation is 120, its ratio to the centred average is 120/102.5, approximately 1.1707. The ratio contains seasonal and irregular effects under a multiplicative interpretation. One ratio alone does not establish the seasonal index for that quarter. Gather comparable ratios across years, average or otherwise robustly summarise them, and normalise the resulting indices. Be careful to preserve the alignment of each average. An off-by-one placement can assign a seasonal effect to the wrong period while leaving every arithmetic operation apparently correct.
7. Estimating a straight trend
A linear trend uses ŷ = a + bt, where t is a time index. Least squares chooses the line that minimises squared deviations from observed values. With time coded so that its sum is zero, a equals the mean of y and b equals Σty/Σt². Centring time simplifies the arithmetic but does not change the fitted trend. For an even number of observations, convenient symmetric codes may use half units or doubled integers. Interpret the slope according to the chosen time scale.
For t values minus one, zero and one with observations ninety, one hundred and one hundred ten, a is one hundred. The cross-product sum is twenty and the squared-time sum is two, so b is ten. At t = 2, the forecast is 120. This extends the fitted historical movement by one period. It assumes the relationship continues and does not account for an unexpected structural change. If time codes advance by two for each calendar year, the coefficient is per coded unit and must be converted before describing an annual change. A unit mistake can double or halve the reported trend even when the fitted observations are correct.
8. Other trend descriptions
A freehand trend is a visual summary and depends on the analyst's judgement. Semi-averages divide a series into two parts, compute each part's mean and connect the means at their appropriate time centres. The method is simple but uses limited information about the detailed pattern and requires a stated convention for an odd number of observations. Moving averages provide a flexible historical trend without imposing a single straight line. Least squares provides an explicit equation and a reproducible criterion.
Growth may be better represented by a percentage than by a fixed absolute increment. An exponential trend can be written Y = A × Bᵗ for positive values. Taking logarithms converts it into a linear relation in log Y. The fitted growth factor is B, and the corresponding rate is B−1. This is an extension for understanding proportional growth, not a reason to fit a curve automatically. Compare the model with the observed pattern and intended forecast horizon. A curve can fit past data closely and extrapolate implausibly. The best historical fit is not necessarily the best forecast. Simpler baselines remain valuable because they reveal whether added complexity produces useful improvement on later observations.
9. Seasonal indices
A multiplicative seasonal index expresses a period's typical level relative to its underlying level. Quarterly indices are commonly normalised to sum to four, or four hundred when expressed as percentages. Monthly indices sum to twelve or twelve hundred. A quarterly factor of 1.10 means a ten percent uplift relative to the underlying level under the model. It does not mean that the quarter grew ten percent compared with the immediately preceding quarter, because the trend and other seasonal factors also matter.
In a simplified stable-level example, quarterly averages are eighty, one hundred, one hundred twenty and one hundred. Their overall average is one hundred, giving factors 0.8, 1.0, 1.2 and 1.0. These sum to four. If the expected deseasonalised level for a future third quarter is 150, the seasonal forecast is 150 × 1.2 = 180. For trending data, raw quarterly averages can confound seasonality with the changing level. A ratio-to-moving-average procedure first estimates the underlying level, calculates observed-to-level ratios, groups them by season and normalises. Additive seasonal effects instead sum to zero over a complete cycle under the usual normalisation. Keep factor and percentage conventions consistent.
10. Deseasonalise and reseasonalise
Deseasonalisation removes the estimated seasonal component to make underlying comparisons easier. Under a multiplicative model divide the observation by its seasonal factor. Under an additive model subtract the seasonal effect. Suppose observed volumes are 96 in a quarter with factor 0.8 and 144 in a quarter with factor 1.2. Both deseasonalise to 120. The difference in observed volumes is consistent with the seasonal pattern rather than a change in underlying level in this illustration.
To forecast, estimate the future underlying level and then reapply the appropriate seasonal component. If the deseasonalised forecast is 130 and the next quarter's factor is 1.2, the final forecast is 156. Using the wrong quarter's factor can create a large error even when the trend calculation is correct. Seasonal patterns can change as customer behaviour, payment systems or holidays change. Review their stability rather than treating an index as permanent. An exceptional festival timing shift may move activity between calendar periods. The model should serve the operational question, and a seasonally adjusted comparison should be clearly labelled so readers do not mistake it for the actual cash or transaction volume the branch must handle.
11. Exponential smoothing
Simple exponential smoothing updates a level estimate using Lₜ = αYₜ + (1−α)Lₜ₋₁, with α between zero and one. The next-period forecast is the updated level in the basic model. A larger α gives more weight to the latest observation and responds faster to changes. A smaller α smooths more strongly and responds more slowly. The method is most appropriate for a series without a persistent trend or seasonal pattern unless additional components are included.
With prior level one hundred, latest observation 120 and α = 0.3, the updated level is 106. If the following observation is ninety, the next update is 0.3 × 90 + 0.7 × 106 = 101.2. The initial level matters particularly early in the series, so state how it is set. Choosing α after inspecting the entire evaluation period leaks future information into the assessment. Select or fit it using training data, then evaluate on later observations. Simple smoothing produces a flat multi-step forecast because it contains no explicit trend. A rising series may therefore be systematically underpredicted. This is a model limitation rather than an arithmetic error in the update formula.
12. Measure forecast error
Define forecast error consistently as actual minus forecast. Mean error reveals average signed bias, but positive and negative errors can cancel. Mean absolute error, or MAE, averages absolute errors and remains in the original units. Mean squared error averages squared errors. Root mean squared error, or RMSE, takes its square root and penalises large misses more strongly than MAE. These measures should be calculated on the same evaluation observations when comparing methods.
For actual values one hundred and 120 with forecasts ninety and 130, errors are ten and minus ten. Mean error is zero, MAE is ten and RMSE is ten. Zero mean error does not mean perfect forecasting. For errors zero and twenty, MAE remains ten but RMSE becomes approximately 14.142. Percentage error measures can aid comparison across scales, but become undefined or unstable when actual values are zero or near zero. A business may also care asymmetrically about errors. Underforecasting cash demand can be more costly than overforecasting by the same amount. Statistical accuracy measures should therefore be interpreted alongside a decision-relevant cost or service measure rather than used as the sole ranking criterion.
13. Evaluate forecasts without hindsight
A realistic evaluation respects chronology. Fit the method using an earlier training period and assess it on a later test period. In rolling-origin evaluation, repeatedly move the forecast date forward and use only information available by that date. Match the horizon to the intended use. A model that forecasts tomorrow well may perform poorly six months ahead. Avoid centred averages, future revised data or parameter choices based on the test period when claiming a real-time forecasting result.
Compare with simple baselines. A naive forecast repeats the latest observation. A seasonal naive forecast repeats the corresponding observation from the previous seasonal cycle. These methods can be difficult to beat for some series and provide a useful reference. Suppose a complex trend model has lower training error but higher test MAE than the seasonal naive method. The historical fit does not justify adopting it. Examine whether a structural change occurred and whether the model's assumptions remain plausible. Report the evaluation dates, horizon, observations and error definition. A forecast is conditional on information and assumptions, not a promise. Planning should allow for uncertainty and for events outside the historical pattern.
14. A branch cash-planning case
A branch wants a next-quarter withdrawal forecast. It checks that the historical figures are quarterly flows in lakh, reconciles missing periods and identifies a recurring third-quarter uplift. The estimated seasonal factor is 1.20. A trend model forecasts a deseasonalised level of 150 lakh, producing a seasonal forecast of 180 lakh. The team compares earlier forecasts with actual withdrawals using chronological test periods and examines whether high-demand misses are operationally costly. The illustrative forecast is a planning input, not a regulatory cash requirement.
If the actual volume is 198 lakh, the error under the actual-minus-forecast convention is eighteen lakh. The percentage error relative to actual is about 9.09 percent. A manager should investigate whether the miss reflects an unusual event, a changed trend or an outdated seasonal factor. Replacing the forecast with the actual after the event would erase the information needed to evaluate the method. Preserve the forecast as issued. Before answering any time-series question, identify frequency, component model, timing and units. Then calculate, check alignment and interpret the result for its horizon. The interactive laboratory lets you vary a trailing window and smoothing weight while keeping the observed series fixed.
Guided exercise: choose the correct time label
Monthly observations are eighty, one hundred, one hundred twenty and one hundred forty. The first trailing three-month average is one hundred and can forecast month four using information through month three. After month four is observed, the next trailing average is one hundred twenty and can forecast month five. A centred three-month average of the first three observations is also one hundred, but it describes month two. Identical arithmetic produces different labels because one calculation smooths history and the other forecasts from a stated origin.
The month-four forecast error from the first trailing average is forty because actual volume is one hundred forty. A straight trend fitted to the first three observations would forecast one hundred forty and happen to be exact in this constructed sequence. That does not establish that a straight trend always outperforms moving averages. The example was designed with constant increments. If month four instead reflected a sudden decline, the comparison would differ. Evaluate methods over an appropriate series of forecast origins and preserve the forecasts actually available at each date. This discipline prevents choosing a method simply because it fits one convenient example. It also explains why a smooth historical chart should not be presented as evidence of accurate real-time prediction without a separate chronological evaluation.
Forecast comparison laboratory
Values are in chronological order. Smoothing starts at the first observation. These methods contain no seasonal adjustment.
Practice and review
25 questions. One answer per question, no negative marking. Explanations remain readable without JavaScript.
Local browser storage only; no account synchronisation.
Ready: 25 questions.
References and scope
- IIBF CAIIB Rules & Syllabus 2026, PDF page 5: Official Module A scope; A01 to A09 are Clerk2Officer editorial identifiers.
- NIST time-series methods: Time dependence, smoothing and trend methods.
Original exposition, cases and questions. Formula conventions are stated in the lesson. Automated checks support arithmetic and functionality; they are not independent academic certification.