{"id":17315,"date":"2013-04-04T01:02:38","date_gmt":"2013-04-04T08:02:38","guid":{"rendered":"http:\/\/drduru.com\/onetwentytwo\/?p=17315"},"modified":"2013-04-04T01:25:48","modified_gmt":"2013-04-04T08:25:48","slug":"t2108-update-130403","status":"publish","type":"post","link":"https:\/\/drduru.com\/onetwentytwo\/2013\/04\/04\/t2108-update-130403\/","title":{"rendered":"T2108 Update (April 3, 2013) – A New Beginnnig for T2108"},"content":{"rendered":"

(T2108 measures the percentage of stocks trading above their respective 40-day moving averages [DMAs]. It helps to identify extremes in market sentiment that are highly likely to reverse. To learn more about it, see my T2108 Resource Page<\/a>. You can follow real-time T2108 commentary<\/a> on twitter using the #T2108 hashtag. T2108-related trades and other trades are posted on twitter using the #120trade hashtag<\/a>)<\/p>\n

T2108 Status<\/strong>: 45.9%
\nVIX Status<\/strong>: 14.2
\nGeneral (Short-term) Trading Call<\/strong>: Hold (1-day trade is to short $SPY based on the T2108 2-Day Decline Model. See below for more details)<\/p>\n

Reference Charts<\/strong> (click for view of last 6 months from Stockcharts.com):
\nS&P 500<\/a> or SPY<\/a>
\nSDS<\/a> (ProShares UltraShort S&P500)
\nU.S. Dollar Index<\/a> (volatility index)
\nVIX<\/a> (volatility index)
\nVXX<\/a> (iPath S&P 500 VIX Short-Term Futures ETN)
\nEWG<\/a> (iShares MSCI Germany Index Fund)
\nCAT<\/a> (Caterpillar)<\/p>\n

Commentary<\/strong>
\nI am FINALLY ready to announce a new beginning for T2108. On Tuesday night, I presented my final project in my data mining and machine learning course (using R as the programming platform), a project based on an analysis of T2108. I am now ready to start rolling out my results and methods. This remains a work in progress as I have already written up nearly a page of things I would like to do, try, and enhance. Thus, I will not yet publish my project until I feel I have addressed several additional modeling questions and done even more thorough data investigations. However, if you would like to see the presentation anyway, I will email it to you if you would kindly agree to send me feedback. (Go to my contact page<\/a> to send me a message.)<\/p>\n

I had one major objective: use data mining techniques to develop a more rigorous definition of oversold and overbought conditions. I produced some encouraging results but not yet as conclusive as I would like. 20% still works OK as a threshold for oversold and 70% is adequate as a threshold for overbought, but other threshold levels hold very interesting properties as well (like 40% on the downside and 75% to the upside) that produce relatively reliable predictions for short-term trades on the S&P 500. Also, the switching behavior I have noted for overbought periods – where the odds switch from the bears to the bulls once the duration hits around 20 days – is a common characteristic for many thresholds. After seeing these characteristics, I decided that the concept of oversold and overbought may be too narrow for how to interpret the crossing of threshold levels. Stay tuned on this. <\/p>\n

A fortuitous objective appeared after the first few weeks of the class. Recall in mid-February I wondered whether a steep 2-day decline in T2108 could be considered a “quasi-oversold” condition<\/a>. I predicted a bounce, and, voila, it happened. That event launched me into an early investigation using data mining to see whether 2-day declines in T2108 could meaningfully predict whether or not the S&P 500 would trade up or down the next day. My preliminary results were encouraging, but incomplete<\/a> (and not quite correct given a very small flaw I found in my data!). I have now developed the model enough to make predictions.<\/p>\n

T2108 declined 17.8% on Tuesday and has declined -18.4% over two days. Using tuned decision trees (classification), my model concludes that Thursday will be a down day for the S&P 500 with a 74% likelihood.<\/strong> Yes, this took my aback as well since I was hoping for more consistent behavior of bouncing after steep declines. But many factors are at work shaping this prediction, and it seems the VIX did not increase enough to generate an up prediction. (The error rate for this model is also better for down predictions than up predictions). Note well that I am not predicting the magnitude of the down day. The different machine learning models I tried were terrible at predicting the magnitude of the decline, and I expected this behavior. The predictors I used in the model do not suggest that they can have anything to say about magnitude. <\/p>\n

The decision tree for this prediction is extremely deep (too deep!), and this is an area where I need to play around with modeling parameters to find the right balance between simplicity of the tree and prediction accuracy (or from a trading perspective, prediction utility<\/em>, that is, the usefulness of the prediction). For example, the decision tree used to make the prediction for Thursday (April 4th’s) trade is as follows (all quantities are based on April 3rd’s closing trade):<\/p>\n