With a very simple trading rule, and a one-day look-ahead model, we show that KDH alternative data generates ∼7%+excess return margin, relative to a baseline model excluding our data product.
We demonstrate leveraging our data for prediction of the S&P 500 Volatility index (VIXD) and UK FTSE All-Share Return Index (TFTASD). Training and testing is conducted using a moving window approach such that we train on features from the day before a given day, and use the given day as the true outcome (see Fig. A). For stability, we ensure that the window is at least 100 days in length.
For prediction, we use a regression tree ensemble (random forest), an off-the-shelf machine-learning tool, with 200 individual trees creating the ensemble.
To measure the informational advantage that our KASPR Global ICT Intel Data product yields, we apply the train–test windowing approach to all days in May 2019 either with both traditional market indicator data and our data product, or with traditional market indicator data only as a benchmark.
After dropping non-trading (weekend) days, we have 21 (FTSE) or 22 days (VIX) of testing days in our sample.
Backtesting Results
Predicting the S&P 500 Volatility index VIX
For the VIX we have a total of 22 trading days in May for testing, using the rolling window train–test methodology, summarised above.
Per the method outlined above, we first train–test with only the traditional indicator data, running 100 independent repeats of the exercise to generate a distribution of outcome accuracy. Then, we follow the identical approach but add the KDH Global ICT Intel data features to the exercise, allowing us to compare accuracy differences against the traditional benchmark.
Results are summarised in Fig. B above, showing that when KDH Global ICT Intel data are added to the feature set (blue) a substantial lift in day-ahead prediction accuracy is achieved over the traditional benchmark (grey). Average accuracy for the traditional features only is 44.3%, whilst with KDH Global ICT Intel data added, average accuracy rises to 59.5%.
Predicting the UK FTSE All-Share Return Index
Next, we apply an identical approach to the FTSE All-Share Return Index, with 21 available days of train–test data available using the rolling window design outlined above.
For the FTSE, we find that average benchmark accuracy of 52.4% is obtained without the KDH Data product measures, whilst average accuracy performance lifts strongly as before with KDH Global ICT Intel data added, to 60.7% (see Fig. C).
Contact us for the complete backtesting report as well as the replication code and data.