What Is Forecasting in Data Mining?

Forecasting is the specific case of prediction where the unknown value being estimated is a future point in a time-ordered series — next month's sales, next quarter's demand, tomorrow's temperature — estimated from historical and current data rather than guessed at from scratch. It's worth being precise about how this relates to prediction more generally, covered in the lesson on classification and prediction: prediction is the broader term for estimating any unknown value from a model, whether that value lies in the future, is simply unobserved (a loan applicant's likelihood of default), or belongs to a different category. Forecasting is prediction applied specifically to a value's position in time. The underlying mechanics of decomposing a time-ordered signal into trend, seasonal, cyclic, and irregular components — the actual machinery most forecasting methods build on — are covered in depth, with a fully worked example, in the lesson on sequence mining; this lesson focuses instead on the reasoning and assumptions that govern how forecasting is used and evaluated.

The Assumptions Every Forecast Quietly Makes

Forecasting isn't just "apply a formula to past data" — it rests on a handful of assumptions worth stating explicitly, because each one is also exactly where a forecast can go wrong.

Past patterns tend to recur. A forecast is only as good as the assumption that whatever generated the historical pattern is still in effect. This is also the assumption's biggest weakness: a genuinely novel disruption — a cyclic downturn, a one-off shock, the kind of irregular component covered in the lesson on sequence mining — breaks exactly this assumption, which is why forecasts built purely on historical repetition tend to fail precisely at the moments they'd be most valuable.

Shorter horizons are more reliable than longer ones. Tomorrow's forecast is more trustworthy than next year's, simply because there's more time for unmodeled factors to accumulate and shift the outcome the further out the forecast reaches.

Aggregated forecasts are more reliable than individual-item forecasts. This one is worth grounding in the actual math rather than taking on faith, because it's the same underlying principle behind why averaging multiple models reduces error (covered in the lesson on bagging vs. boosting) and why larger samples reduce estimation error (covered in the lesson on numerosity reduction). If a single product's weekly demand fluctuates with a relative noise level of 40% around its average, and a company sells 20 roughly independent products of similar volume, the total demand's relative noise level isn't 40% — it shrinks by a factor of the square root of the number of products being summed:

relative noise of the total = 40% / sqrt(20) = 40% / 4.47 ≈ 8.9%

Individual product-level noise partially cancels out when many independent products are summed together, the same way individual tree errors partially cancel when averaged in a bagged ensemble. This is precisely why forecasting total company revenue is a fundamentally easier and more reliable problem than forecasting any single product's sales.

A forecast is never exactly right. Because of the three points above, a forecast is more honestly expressed as a range (say, 90,000–110,000 units) than as a single number — the range communicates the genuine uncertainty a single point estimate hides.

Qualitative vs. Quantitative Forecasting

Forecasting methods split into two families based on what they're built from. Quantitative forecasting uses historical data and mathematical or statistical models — time series analysis, causal/regression models — and is generally more consistent and reproducible, precisely because it doesn't rely on any individual's judgment. Qualitative forecasting relies on expert opinion and experience rather than historical data, and is what's actually available when little or no relevant historical data exists yet — launching a genuinely new product with no sales history, for instance. It's important to place this correctly: qualitative forecasting is still forecasting, just the less rigorous end of the same spectrum, not some fundamentally different, unscientific activity sitting outside forecasting altogether. Its acknowledged weakness is exactly what you'd expect from relying on judgment — it's more exposed to personal bias than a model built on historical data.

Types of Forecasting

• Economic forecasting predicts broad economic indicators — inflation, GDP growth, unemployment — typically used by policymakers and investors.
• Technological forecasting projects future technology trends, helping organizations prepare for or invest ahead of coming shifts.
• Demand forecasting estimates future customer demand for a specific product, directly feeding sales, production, and marketing planning — this is the type most directly built on the techniques covered throughout this series.

What Makes a Forecast Good

A forecast earns trust along a few concrete dimensions: accuracy (how close it lands to the eventual real outcome), reliability (consistent performance across repeated forecasts, not just accurate once by chance), timeliness (available with enough lead time to actually act on), clarity (understandable to the people who have to use it, not just the person who built it), and cost-effectiveness (the value of better decisions it enables has to exceed what it costs to produce).

Choosing a Forecasting Technique

Two practical questions settle most of the choice: does the situation call for a single point estimate or a range, and how far into the future does the forecast need to reach? The time-horizon question matters most directly, since it interacts with the reliability assumption above — a technique that performs well forecasting next week's demand may be substantially less trustworthy stretched out to forecast next year's, regardless of how sophisticated the underlying model is.