Lab · C · Change detection
The calibration of the detector that answers the question "is something happening right now", and its evaluation on real data.
Synthetic calibration
What it is: With what delay the detector catches a shift in artificial series where the date of the shift is known, and how often it raises a false alarm when there is no shift.
What data feeds it: Produced only with synthetic data; it does not touch the database. Each scenario is repeated 60 times; the first 90 observations are the warm-up period.
What it shows: The average detection delay (days), the average time until the first false alarm (ARL0) and the number of scenarios caught; separate rows for mean shifts and volatility increases.
What can be concluded from it: The detection delay is inversely proportional to the square of the shift-to-noise ratio: small mean shifts are almost never caught, while a doubling of volatility is caught within a few days. This is why the detector monitors volume and volatility.
Alarm timing around events
What it is: The distribution of alarms in the ±10 days around recorded events, and a comparison with random days.
What data feeds it: The event table and the alarm series of the selected run. The same statistic on random dates is computed as the null hypothesis.
What it shows: Days before the event are amber bars, days after it grey bars. Below, the pre-event alarm rate and the random rate.
What can be concluded from it: If the pre-event rate is not clearly higher than random, the detector does not sense events in advance. If there are no event records, the chart is empty; earnings and KAP dates should be loaded from Settings > Events.
Persistence after an alarm
What it is: Whether the sign of the direction on the alarm day is preserved 1, 5 and 21 days later.
What data feeds it: The alarms of the selected run and the residual returns in return_daily.
What it shows: Hit rate per horizon, 95% interval and number of observations.
What can be concluded from it: If the rate is below 50%, the move on the alarm day reverses afterwards. Noticing a change early does not mean profiting from it; on real data this rate comes out at 47–48%.
Daily alarm count and market turbulence
What it is: How many stocks raised an alarm each day, and the probability that the market is in a turbulent regime.
What data feeds it: The alarm series of the selected run, and the filtered probability of a two-state Markov regime model fitted to the XU100 return.
What it shows: Amber bars are the alarm count, the red line the turbulence probability. The title shows the overall alarm rate and whether neighbour evidence is on.
What can be concluded from it: Alarms piling up in turbulent periods show that the detector is catching market-wide rather than stock-specific events. The filtered probability is used, not the smoothed one: the smoothed series sees the future and is therefore invalid in live use.
The investment information, comments and recommendations given here are not within the scope of investment advisory services. Investment advisory services are provided under an investment advisory agreement to be signed between a client and brokerage houses, portfolio management companies, or banks that do not accept deposits. The signals here are produced from historical data with statistical models, are shown the same to everyone and are not personalised; they may not suit your financial situation or your risk and return preferences. Therefore, making investment decisions based solely on the information given here may not produce results that meet your expectations.