LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A time-series forecast is an expectation under the pattern already in the data. It is not a call on a break that the series has not shown. Filtering estimates the current state from the past and the present. Smoothing estimates a past state using the whole path. Both remove a defined noise. Neither knows about a torn muscle, an earnings jump, or a mortgage book until that information is in the series or brought in from outside it.
Why it matters
Classroom work often assumes stationarity: mean, variance, and autocovariance follow a fixed rule. Firms do not. A twenty-year-old company is not the same company, and there is no one-shot test that settles it. Prices are often treated as having a unit root, so people model returns and then argue. A no-arbitrage assumption is mild, because a static mispricing is traded quickly. Constant volatility is a bad description of a whole surface and still useful as the definition of implied volatility. A wrong forecast can still be a usable level.
The idea in plain words
A regular series uses a fixed clock, such as a daily close. An illiquid name has uneven gaps. Joining the dots hides what happened between prints. Under the chart is a firm or a credit book. The series is a consequence.
A decomposition, when the pieces exist, splits trend, seasonality, and shock. Trend is the long drift. Seasonality repeats on a calendar. The shock is the noise the model assumed, sometimes with a moving variance. Stable trend and season make the forward expectation look like the recent path. A new product can kill a sales trend. The shock term does not automatically include that break.
A moving average filters: the window estimates the present, with a band around that fit. It does not speak about the future. Exponential smoothing uses later points to describe earlier ones, so it needs the whole path. The number that minimizes squared error for a fair die is 3.5, which a die cannot show. That is why the lecture refuses the word prediction.
How to use it
Three breaks make the limit obvious. An earnings gap jumps the price to a new level, and a naive forecast stays at the old one. The lecture would rather express that event with the option market: a straddle for the size of the uncertainty and the market's lean on the direction. The straddle does not tell you the direction without information the series does not contain. December temperature in New York is the intuition. You do not forecast 100 degrees Fahrenheit. If yesterday was about 20, tomorrow's expectation is near 20. The model is organizing a pattern a person already understands.
A bench log from January 2024 to January 2025 sits near 350 pounds, then on 9 December 2024 prints 45 pounds after a torn pectoral muscle. A doctor knew before the chart did. Parking-lot photos, a CEO's face, or mortgage tapes are the same kind of outside information. In 2008 a continuation forecast of credit-default swaps keeps rising while the realized path collapses. People who read the loan tapes were using information the series had not priced. That episode is a pattern, not a template to rerun.
Some jobs do not need the outside information. A dice market quoted around 3.5 collects the spread because the mean does not move. Let the mean drift and a moving average can still collect a spread while being wrong about the exact level. An illiquid price, seen only as scattered trades, can be filled in from liquid relatives with a filter in the neighborhood of a Kalman filter. At high frequency a jump may be something you re-center on.
What usually goes wrong
Calling the output a prediction, then being shocked when a break makes it absurd, is the main miss. Assuming the classroom moments are constant is the second. Ignoring a unit-root argument and forecasting the price level through an earnings gap is the third. The fourth is the opposite error: throwing the filter out of a market-making or illiquid-pricing problem where a wrong-but-centered level is the object you quote around. Alternative data is not mandatory in those two settings. It is mandatory when the question is a structural break the price has not printed yet.
Questions people ask
What is the difference between filtering and smoothing?
Filtering uses past and present data to estimate the current state. Smoothing uses the whole path, including later points, to estimate what a past state was. A forecast pushes an estimate forward.
Why is 3.5 the best dice call?
It minimizes squared error across rolls. It is not a face the die can show. The lecture uses that fact to separate an expectation from a prediction.
When did the bench series break?
On 9 December 2024 the load dropped from a cloud around 350 pounds to 45 pounds. The cause was an injury, which was known to people in the room before it was a point on the chart.