Time Series and Text Data

4.1 Time-Series Foundations

Unit 3 regression uses cross-sectional data—different objects observed at one point in time. This unit uses a different structure: observations of the same object at different points in time. Forecasting separates a pattern from history and projects it one step into the future.

Time-Series Data and Cross-Sectional Data

Four Components of a Time Series

Time-Series Models versus Regression

Example: Six Months of Fastchips Sales

What should be the forecast for month 7?Different methods produce different answers—the naïve method forecasts 28, the average method forecasts 25, and a moving average gives yet another value. The method should match the stability of the data.

Six Forecasting Methods and Their Applicability

Measuring Forecast Error: MAD and MSE

• MAD(mean absolute deviation) is the average absolute forecast error and measures the typical size of error.
• MSE(mean squared error) is the average squared forecast error and penalises large errors more heavily.
• Both measures depend on units, so comparisons must use the same series and the same evaluation interval.
• Error is defined as actual − forecast, which has both positive and negative values; take the absolute value or square it before computing MAD and MSE.

4.2 The CCW Case and Seasonal Factors

Case: Forecasting Call Volume at CCW

CCW sells low-cost computers directly by telephone and operates its call centre around the clock. Lines are limited and new customers cannot get through when all lines are busy, so staffing must be planned in advance and call volume must be forecast. The data are daily average call volume in each quarter for the past three years.

CCW’s Current Practice: The 25% Rule

Call volume is relatively stable except around Christmas. The company assumes each quarter is the same and only Q4 is 25% above the other quarters:

• Forecast for next year’s Q1 = actual Q4 ÷ 1.25
• Forecasts for Q2 and Q3 = previous quarter’s actual value
• Forecast for Q4 = 1.25 × Q3 actual value

Using Q5–Q12 as the comparison interval, the 25% rule’s MAD = 519.9, MSE = 420,476.

Seasonal Factors and Deseasonalisation

• Seasonal factor = quarterly average call volume ÷ overall mean, measuring how far the quarter is above or below the annual average.
• The four seasonal factors sum to 4.
• Deseasonalised value = actual value ÷ seasonal factor for that quarter. Removing seasonal variation makes the trend easier to see.

Three Steps for Forecasting Call Volume

1. Choose a time-series forecasting method.
2. Apply it to the deseasonalised series to obtain a deseasonalised forecast.
3. Multiply the forecast back by the seasonal factor for the corresponding quarter to obtain the final forecast.

4.3 How Methods Perform on the CCW Data

This section answers one question: once seasonality is handled, how much do simpler forecasting methods actually differ? The answer may surprise you.

Conclusion: The Method Is Not the Main Issue

• The current 25% rule has MAD = 519.9, the largest error among the six approaches.
• After seasonal adjustment, the naïve method reduces MAD to 272.4.
• Once seasonality is removed, the simple methods perform similarly.
• Recognising seasonality matters far more than choosing one smoothing method over another.
• MAD within the exponential-smoothing family changes with initialisation and parameter choices. This page consistently uses F₁ = first-period actual value and α = 0.3; the Excel template initialises slightly differently, causing differences of 1%–15%. Keep the method consistent when comparing.

4.4 Causal Forecasting and Judgemental Forecasting

What Causal Forecasting Is

Time-series forecasting uses only the series’ own history;causal forecasting finds and validates a relationship between independent and dependent variables to forecast future Y. Unlike time-series forecasting, it requires an interpretable driver variable, and that driver must itself be forecast. Unit 3 regression is the tool for causal forecasting.

Judgemental Forecasting Methods

These methods are especially useful when data are scarce, a new product is launching, or market structure is changing. Their risk is mistaking personal preference for fact, so cross-check them with quantitative methods.

The Consultant’s Advice to CCW

• Switch forecasting to monthly data, and make hiring and training decisions monthly to avoid large one-off adjustments.
• For established products, begin with a low smoothing constant in exponential smoothing—demand is stable, so do not overreact to noise.
• For new products, use a high smoothing constant in exponential smoothing with trend—growth-stage conditions change quickly and the forecast must keep up.
• All methods should first deseasonalised.
• forecast by segment and then aggregate, finally using sales regression to obtain call volume.

4.5 Text Data and AI Labelling

This section treats AI as a labelling machine: a large language model converts unstructured reviews into sentiment and topic labels, and human sampling plus an agreement rate determine whether the labels can be used.

Questions That Text Data Can Answer

• What customers complain about and praise in reviews.
• How word-of-mouth differs across products.
• How topic distributions change over time.
• Text is unstructured data and cannot enter regression directly—first convert it into countable labels.

How LLMs Process Text—and Why They Err

The model splits text into tokens and predicts the probability of each next token. For labelling, it is essentially continuing an answer that follows the required format, so the output looks like a classification even though it remains probabilistic generation:the same question asked twice may produce different labels.

Six Steps in Text Analysis

Human Sampling and Agreement Rates

Randomly select 20 labelled records and judge them manually again. Compare whether the AI and human labels match for each record:agreement rate = matching records ÷ sampled records. The key issue is not simply the level of agreement, but which topics account for most disagreement—clusters of disagreement indicate unclear definitions or a model weakness.

Limitations and Pitfalls in Text Analysis

• Sarcasm and negation are easily misclassified (“Great, it arrived broken”).
• Industry jargon and abbreviations must be defined in the prompt.
• A long review may contain more than one topic.
• Asking twice about the same data may change the conclusion.
• Thousands of texts exceed chat-mode limits and require an API or programming.

4.6 Summary and Assignment

• Time-series forecasting first separates trend, seasonal and random components, then extrapolates.
• After seasonal adjustment, the CCW case’s error can fall from 519.9 to 272.4.
• No forecasting method is universally best; choose based on data stability, and recognising seasonality matters more than the particular method.
• Causal forecasting requires an interpretable driver variable, and that driver must itself be forecast.
• The new text-analysis workflow is AI labelling + human sampling + agreement rate and requires no coding; agreement and stability determine whether the results are usable.

Appendix: Interactive Demonstrations and Classroom Exercises

Data sources: CCW.xlsx (12 quarters of daily average call volume), 单元4_时间序列课堂练习.xlsx (12-period sales series), Table 10.41 in Forecasting.xlsx (sales and call volume), and 单元4_评论标注练习_中文.xlsx (100 reviews and 20 sampled records). The review text is simulated course data, not real-platform user reviews.

Unit Glossary (English–Chinese)