FIN40090 Stacked Ensembling and Algorithmic Trading Project (Group) The purpose of this project is to bring together the methods we have had so far this semester to produce a forecast of (‘trimmed’) daily return of a relatively large Equity or major Equity Index, and to carry out a study of simulated trading which includes at least 250 days of ‘out-of-sample’ performance. The study should be documented and presented in the form of a report (single pdf file, no longer than 15 pages, including printout and tables) and a summary presentation of 5 minutes (plus time for questions) to be delivered jointly by group members in the penultimate class [the presentation part may be pre-recorded]. Specifically, the steps included are as follows: 1. Agree (and submit) one choice of Equity and one of Equity Index (other than FTSE) for your group to work on. You will be notified which of these your group will be assigned to work on. 2. Prepare input and output data matrices of lagged daily (‘trimmed’) logarithmic return, removing non-trading (exactly zero-return output) days as shown in class. [In addition to lagged return information, you should include several lags of log(1+Volume) as Inputs] 3. Next, for each of the following methods and models, produce ‘out-of-sample’ forecasts using either ‘5-fold’ or (ideally) ‘Out-of-Bag’: Linear Regression (w/Ridge or Lasso), Logistic Regression (‘Up’/’Down’), Support Vector Machine, Gradient Boosting Other method of your choice (preferably from HW) Each group member should take the lead on one of the methods above, which should be clearly indicated. 4. Include the ‘Expert’ forecasts from the above alongside the original input variables and run a Deep (3 or more hidden layer) Neural Network to forecast return from the combined inputs. Use the (m(x)y-1)^2 loss function as discussed in class. 5. Perform an analysis of trading performance of the final Stacked Ensemble model using (a) ‘Long-Short’ investing and (b) ‘Proportional’ (to forecast) investing, and present cumulative P/L plots and (annualized) Sharpe Ratios, highlighting the performance on (‘out-of-sample’) ‘testing’ data.

学霸联盟