Time Series Basics - Business Analytics

1. What is the most appropriate conclusion about this proposed decomposition?
   Answer: One seasonal period cannot directly represent both cycles, so another seasonal effect should be modeled separately When you encounter a time series with multiple repeating cycles at different frequencies, your first instinct should be to ask: can a single seasonal model capture all of them? Classical decomposition assumes exactly one seasonal period — you pick a value like $s = 7$ for weekly or $s = 365$ for annual, and the model extracts one repeating pattern of that length. Here, the data has two distinct cycles: a weekly rhythm ($s = 7$) and an annual rhythm ($s = 365$). These operate at completely different frequencies and cannot be collapsed into one period. That's precisely why C is correct — one seasonal period is structurally incapable of simultaneously representing both a 7-day and a 365-day cycle. The right solution is to model the second seasonality separately, using frameworks like STL decomposition or TBATS, which explicitly handle multiple seasonal periods. A is wrong because annual peaks don't consistently land on the same weekday each year — the calendar shifts, so weekly and annual cycles are not synchronized and cannot be merged into a single period of $7$. B is wrong because weekly variation does not average away into a stable trend; it persists as a real, recurring pattern that an annual period would misattribute or ignore entirely. D is wrong because detrending and seasonal decomposition are sequential, independent steps — removing the trend doesn't cause two distinct seasonal frequencies to merge into one. As a study tip: whenever a question describes a time series with cycles at two different lengths (daily vs. yearly, weekly vs. monthly, etc.), that's your signal that classical single-period decomposition is insufficient — you need a multi-seasonal approach.
2. Which decomposition choice is best supported by this pattern?
   Answer: Use a multiplicative decomposition because seasonal amplitude rises roughly with the series level When choosing between additive and multiplicative decomposition, the key diagnostic is how seasonal variation behaves relative to the series level. Ask yourself: does the seasonal swing stay roughly constant in dollar terms (additive), or does it grow and shrink proportionally with the overall level (multiplicative)? The passage tells you exactly what you need: as average monthly revenue grew substantially over six years, the peak-to-low seasonal gap doubled from $\$20K$ to $\$40K$. Crucially, this difference "remained approximately constant as a percentage of the series level." That proportional stability is the textbook signature of multiplicative seasonality — the seasonal component scales with the trend, so modeling it as a ratio (rather than a fixed dollar amount) captures the pattern correctly. Answer B is right because it identifies this proportionality as the deciding evidence. A is wrong because it confuses measurement units with decomposition logic. Yes, you can always measure differences in dollars, but that doesn't make the pattern additive. Additive decomposition requires that the seasonal amplitude be roughly constant in dollars — which it isn't here; it doubled. C introduces a false prerequisite. Multiplicative decomposition is appropriate whenever seasonal amplitude is proportional to the level, regardless of whether the underlying trend follows an exponential curve or any other specific functional form. D is a trap for students who notice that the dollar magnitude changes and mistakenly conclude seasonality is irregular or should be dropped. Changing dollar amplitude with a stable percentage relationship is exactly what multiplicative seasonality looks like — it belongs in the model, not outside it. A practical rule: if seasonal swings grow with the series, go multiplicative. If they stay flat in absolute terms, go additive.
3. Which conclusion follows from correctly deseasonalizing both observations?
   Answer: Beta has the stronger underlying level because its deseasonalized sales are $1{,}200$ units Whenever you see a question involving seasonal indices, your first move should always be to deseasonalize the data before comparing stores, regions, or time periods — observed sales alone are misleading because they reflect both the true underlying demand and the seasonal boost or drag. To deseasonalize, divide observed sales by the seasonal index: $\text{Deseasonalized Sales} = \frac{\text{Observed Sales}}{\text{Seasonal Index}}$ For Store Alpha: $\frac{1{,}200}{1.20} = 1{,}000$ units. For Store Beta: $\frac{960}{0.80} = 1{,}200$ units. Even though Alpha's raw sales look higher, Beta was operating in a weak season (index below 1.0), meaning its baseline demand is actually stronger. Beta's deseasonalized level of $1{,}200$ units is the correct conclusion — confirming B is right. A is the classic trap: comparing raw observed sales without removing the seasonal effect. Alpha's 240-unit advantage in observed sales is entirely explained by its favorable season, not superior underlying demand. C sounds sophisticated but is wrong — the indices don't "cancel out" the difference; they reveal it. The whole point of deseasonalization is that the indices expose the true difference rather than neutralize it. D miscalculates Alpha's deseasonalized figure by multiplying instead of dividing: $1{,}200 \times 1.20 = 1{,}440$, which would actually add more seasonal inflation rather than remove it. A reliable rule: when seasonal indices are involved, always divide to deseasonalize, and never compare stores using raw figures.
4. What should the analyst do before using these values to estimate monthly seasonal effects?
   Answer: Average each pair of adjacent moving averages so the trend-cycle estimates align with individual months When working with classical decomposition, you need to align your trend-cycle estimates with specific time periods before you can isolate seasonal effects. The challenge with a $12$-month moving average is that each calculated value lands between months — for example, a moving average centered on months 1–12 falls between June and July, not on any actual observation. You cannot subtract a trend-cycle estimate from a data point unless both correspond to the same calendar month. The fix is centered moving average (CMA): average each consecutive pair of adjacent $12$-month moving averages. This two-step process — first a $12$-period MA, then a $2$-period MA of those results — produces a **$2\times12$ moving average that lands precisely on individual months. Once your trend-cycle estimates align with real months, you can compute seasonal-irregular ratios (or differences) and derive monthly seasonal indices. Choice A** describes exactly this standard procedure. Choice B is tempting but wrong — arbitrarily assigning each moving average to the sixth month ignores that the value is centered between months 6 and 7, introducing a timing error that distorts every seasonal estimate. Choice C misidentifies the solution: a second $6$-month moving average would create a $6\times6$ structure, not the $2\times12$ needed to center on individual months. Choice D — subtracting the overall series mean — is irrelevant to the alignment problem and is not a step in classical decomposition at all. Study tip: Whenever you see an even-numbered moving average in decomposition questions, immediately think "centering required." The phrase "$2\times12$ moving average" is the exam signal that this alignment step has been correctly applied.
5. Which assessment best accounts for trend and seasonality?
   Answer: The decline from December may be seasonal, while the January comparison is consistent with underlying growth When analyzing demand data, you must separate two distinct forces: trend (the long-run direction of demand) and seasonality (predictable fluctuations tied to the calendar). A common trap is treating any single drop or rise as a trend without asking why that change occurred. Here, you have three data points: January last year ($100$), December last year ($140$), and January this year ($110$). The right move is to compare like periods — January to January — for trend, and recognize that the December-to-January drop likely reflects seasonality, not a reversal of growth. Since January is described as a low-demand month and December as a high-demand month, the $140 \to 110$ decline is almost certainly a seasonal pattern. Meanwhile, January grew from $100$ to $110$, suggesting underlying positive demand growth. This is exactly what C captures — the December decline is plausibly seasonal, while the January-over-January comparison points to real growth. Choice A makes the classic mistake of comparing across different seasons (December to January), treating a seasonal dip as a negative trend. Choice B takes the January improvement and mechanically labels it a "$10$ percent monthly trend" — but two annual observations separated by twelve months cannot establish a monthly trend rate. Choice D is simply false; seasonal analysis does not require daily data. Monthly or even quarterly observations are routinely used for seasonality decomposition. As a study tip: whenever a question mixes months with known seasonal patterns, always compare same-period observations for trend and treat cross-season comparisons with skepticism. Conflating seasonality with trend is one of the most common errors in demand forecasting questions.
6. Which revision would produce the most defensible estimate of future forecasting performance?
   Answer: Estimate decomposition components using only the training period, then project trend and apply seasonal factors to validation months Whenever you see a question about forecast validation, the core principle to keep in mind is data leakage: information from the validation period must never influence how the model is built, or your accuracy estimates will be artificially optimistic and meaningless for predicting real-world performance. Here, the analyst decomposes the entire series — including the final 12 months — before evaluating those same months. This means the seasonal and trend components are informed by the very data being "tested," which contaminates the validation. The most defensible fix is D: estimate all decomposition components (trend, seasonality, residuals) using only the training period, then project the trend forward and apply those training-derived seasonal factors to the validation months. This mimics what would actually happen in production — you build your model on past data and forecast into an unknown future. A is too extreme. Decomposition can use out-of-sample data in exploratory analysis, but the problem here is specifically about forecasting performance estimation, not decomposition as a method. Eliminating the validation period destroys your ability to evaluate the model at all. B commits the exact error the question is testing. Calling seasonal estimation "descriptive rather than predictive" is a rationalization — when those seasonal factors shape your forecasts, they are functioning predictively, and using future data to estimate them is still leakage. C reverses the logic entirely. Estimating seasonality from validation data and errors from training data produces no coherent or valid evaluation framework. Study tip: Anytime a question involves validation, ask yourself: "Did any validation-period information touch the model-building step?" If yes, the methodology is flawed — that's almost always the trap being set.
7. What are the missing fourth-quarter seasonal index and the corresponding observed-sales forecast?
   Answer: An index of $1.15$ and a forecast of $230$ units Whenever you see a multiplicative decomposition problem, your anchor concept is the normalization constraint: in a quarterly model, the four seasonal indices must average exactly $1.00$, meaning they must **sum to $4.00$**. Start by finding the missing fourth-quarter index. You know the first three indices sum to $0.80 + 0.95 + 1.10 = 2.85$. To reach the required total of $4.00$, the fourth-quarter index must be $4.00 - 2.85 = 1.15$. This eliminates B (index of $1.05$) and C (index of $0.85$) immediately — both fail the normalization check. Now apply the multiplicative model to forecast sales. The formula is: $\hat{Y} = T \times S \times I$ where $T$ is the trend-cycle value, $S$ is the seasonal index, and $I$ is the irregular factor. Plugging in: $200 \times 1.15 \times 1.00 = 230$ units. That confirms D as correct. A is tempting because it gets the index right ($1.15$) but then reports $200$ units as the forecast — a critical arithmetic error. Stating the trend-cycle value alone ignores the entire point of seasonal adjustment: the index scales the trend. B and C both use incorrect indices derived from faulty reasoning (perhaps dividing rather than subtracting, or misremembering the constraint). Study tip: Always write out the two-step process separately — find the missing index via the sum-to-$4.00$ rule, then multiply through the full model. Students who try to do both steps mentally often apply the right index but forget to multiply, landing on answer A.
8. How should the analyst interpret the large positive residual?
   Answer: It is an unexplained positive deviation that warrants investigation, but decomposition alone does not establish its cause Whenever you see a question about time-series decomposition and residuals, remind yourself what the residual actually represents: what's left over after the model has accounted for trend and seasonality. That leftover captures irregular, one-off variation — but capturing something is not the same as explaining it. A large positive residual during the week a competitor closed stores is exactly what you'd expect to see if that event boosted sales abnormally. The residual flags that something unusual happened, and the timing is certainly suspicious. However, decomposition is a statistical partitioning method, not a causal analysis tool. Correlation in timing does not establish causation. The analyst should treat the residual as a signal worth investigating further — through additional analysis, interviews, or external data — not as proof of anything. That makes A the correct interpretation. B is wrong because a seasonal index, by definition, must be recurring on a predictable calendar cycle. A one-time competitor closure is an irregular event, not a repeating pattern. Forcing it into a seasonal component would corrupt the model. C commits the classic post hoc reasoning fallacy. Removing systematic variation narrows the field of suspects but does not prove causality. Other coincidental factors (a local event, a promotion, weather) could also explain the spike. D is wrong because a single irregular residual should not trigger a permanent upward revision to the trend. The trend reflects long-run direction; absorbing a one-off shock into it would distort all future forecasts. Study tip: On residual-interpretation questions, always distinguish between detection (what the residual tells you) and diagnosis (what additional investigation is needed). Decomposition does the former, not the latter.
9. What seasonal component is implied for that month?
   Answer: $35$ thousand units, because the negative irregular component must also be removed Whenever you see additive decomposition, the key is remembering that the model equation is a strict accounting identity: every component must sum exactly to the observed value. Given $Y_t = T_t + S_t + R_t$, you can always isolate any missing component algebraically — just rearrange. Here, you're given $Y_t = 248$, $T_t = 220$, and $R_t = -7$. Solving for $S_t$: $S_t = Y_t - T_t - R_t = 248 - 220 - (-7) = 248 - 220 + 7 = 35$ The seasonal component is 35 thousand units, confirming answer A. Subtracting a negative irregular term means adding it back — the irregular component dragged sales down by 7, so without it, the seasonal contribution must account for that full gap plus the excess over trend. Answer B (28) is the most tempting trap. It calculates $248 - 220 = 28$, which is simply the raw deviation from trend — it ignores the irregular component entirely, as if $R_t = 0$. Answer C (21) compounds the error in the opposite direction: it subtracts the irregular component instead of accounting for its sign correctly, computing $28 - 7 = 21$. This treats a negative irregular as something that reduces the seasonal estimate rather than understanding the algebra. Answer D (−35) incorrectly makes the seasonal component negative, perhaps by misreading the sign conventions or confusing which direction the components offset each other. Study tip: Always go back to the equation. In additive decomposition questions, simply isolate the unknown — don't interpret intuitively until you've done the arithmetic first, since sign errors are the most common trap.
10. What is the appropriate point forecast for observed subscriptions in that quarter?
    Answer: $5{,}900$ subscriptions, obtained by applying the fourth-quarter seasonal multiplier Whenever you see a multiplicative time-series decomposition, your job is to reassemble the components — not add them, not pick one over another. The model $Y_t = T_t \cdot S_t \cdot R_t$ tells you that the observed value is the product of trend-cycle, seasonal index, and irregular factor. Here, the calculation is straightforward: $Y_t = 5{,}000 \times 1.18 \times 1.00 = 5{,}900$. The seasonal index of $1.18$ means the fourth quarter typically runs 18% above the trend level, so multiplying inflates the forecast appropriately. Since the irregular factor is $1.00$ (no random disturbance expected), it drops out of the product. Answer C correctly applies this logic. Answer A is a trap for students who confuse forecasting with seasonal adjustment. Dividing by the seasonal index ($5{,}000 / 1.18 \approx 4{,}237$) removes the seasonal effect — useful for analyzing underlying trends, but the exact opposite of what you need when producing a forward-looking forecast of observed subscriptions. Answer B confuses the multiplicative model with an additive one. Adding the index to the trend ($5{,}000 + 18 = 5{,}018$) only makes sense if the model were $Y_t = T_t + S_t + R_t$; in a multiplicative framework, indices are multipliers, not increments. Answer D has no methodological basis — rounding to $6{,}000$ arbitrarily discards precision with no analytical justification. Study tip: Always identify the decomposition type first (additive vs. multiplicative). In a multiplicative model, seasonal indices are ratios — they multiply; they never add.

Start this quiz