TSCI.pngLag Measurements


The lag computations in the tradescape platform are done as follows:

1. An EM signal is computed that has the same count of trades as the user defined trading signal being analyzed.

2. A correlation lag is estimated between the unlagged EM signal and user-defined signal. This is reported as the CorrLag in the report function and as the Lag: All Data in the hint flyovers.

3. The user signal is then mapped point by point to the reference (EM) signal. For each turn in the EM signal, an algorithm searches within a given lag tolerance for a corresponding turn in the user-defined trading signal. If one is found, it is stored in a table of upside, downside, and all turns with the lag measured in each. Only those turns that coincide are accounted. This yields the At Turns, Entry, and Exit Lags. These are robust averages that use the interior 68 percentile of the lags accumulated. Also reported as the uncertainty () is a robust SD that represents the size of the lag at these percentile bounds.

4. The coverage error is reported. It is the % of the EM signals turns which were not covered by the real-world trading signal. This value is ideally less than 10% if trading order, or perhaps 15-25% if trading a combination of order and chaos.

Note: you must be careful of signals that contain massive whipsaws.

An example would be a differenced single pass moving average whose smoothness is insufficient to manage the inherent noise in the differencing procedure. In such instances, the correlation peak can occur not at the lag of the signal, but rather at a point representative of the much faster whipsaws.

excel02.png

This is an extract from two nearly identical signals, both based on differenced 1-pass 80-length SMAs. We know the theoretical lag should be (n-1)/2 or 39. The second case has a correlation lag of 40.07. The first case reports 11.49. In the first case, the algorithm has been fooled by the faster oscillations and since the lag computations at the turns use this correlation lag to establish initial bounds, these are incorrect as well.

The key point to look for is the coverage error of the turns. With a trade to enter and another to exit, keep in mind that a 50% error nominally represents that point where no round-trip trades actually match.

If the coverage error is excessive, such as 40% or higher, you should assume the lag computations are not likely to be accurate, and they may be fully invalid. In the above extract, note that both coverage errors are greater than 50%.

If a signal generates more than a coverage error of 30% or higher, it will be shown with an ! following the signal number (ie., 5!) in the point in the tradescape plot. When you see a signal shown as such, you should be very cautious of the results from the lag analysis, and doubly cautious of using this signal for trading the order in the price movements.