What you will learn
- Clustering
- Jumps
- Intraday seasonality
- Longer-horizon aggregation
Clustering
Volatility clustering describes periods in which large or small movements tend to occur near similar movements. It does not mean the direction of the next return is known. A magnitude pattern is different from a directional signal.
Jumps
Jumps are abrupt changes that a smooth variation model may handle poorly. Their identification depends on sampling and a specified method. Include event and data-quality investigation before assigning every large observation to a market jump.
Intraday seasonality
Intraday seasonality can affect typical activity and spread. Compare like times and sessions when estimating a baseline. A time-of-day pattern can change and should be evaluated across dates rather than assumed permanent.
Longer-horizon aggregation
Longer-horizon risk aggregates more than a simple count of short intervals when dependence, changing volatility or gaps matter. Stress scenarios complement a point forecast but do not establish a complete worst-case bound.
Worked example
A model trained on quiet hours systematically underestimates movement around scheduled releases. The problem may be a mismatched conditional baseline, not merely a need to multiply every forecast by the same constant.
Try it yourself
Design a diagnostic separating ordinary-session errors from event-window errors while avoiding definitions chosen only after the worst outcomes.
Show the worked solution
Predefine event windows from the calendar, compare forecast errors in each group and report sample sizes. Keep a later evaluation period for any revised model and preserve the original errors.
Apply this to your course project
Compare several volatility estimates and document where their forecasts fail.
Keep the calculation inputs, assumptions and decisions with your work. Practical exercises are self-reviewed; the scored knowledge checks assess the questions shown, not an independent certification of practical competence.