Thread: userque's Account talk

Previously used two IFT's this month already.

COB 4-24-14, Exiting F-Fund and moving to G-Fund 100%.

In S fund COB today, one IFT left.

For my personal purposes, I expect to develop a artificial neural network (NN) to predict before noon (11:45am EST or so) whether the close will be "up" or "down."

All of my trading bots/NN's use data dating back to at least 2007, this "DEADLINE-NN" will only start out knowing the last 30 days of trading history. Each new trading day will add to this database.

It will use intraday data--hard to find for free on the internet. If I found it for free, I would feel better about taking the time to code a spreadsheet to extract the data from it that I need. I'm taking the lazy route as this NN is not critical for me and I don't want to pay for the data right now without conducting this proof of concept.

All that to say: If there is sufficient interest, I'll post the predictions here, if and once the NN starts hunting.

That's cool. I'm interested. Thanks!

Originally Posted by sdouglas3

That's cool. I'm interested. Thanks!

OK, we've reached sufficient interest

I'll give weekly updates on the NN's progress until it starts hunting "good." In a best-case scenario, it'll be hunting "good" by next week. If not, I'll post my most recent findings and prognosis starting next week.

If and until they let us IFT near the closing bell, this is the best alternative, imo.

Thanks.

Entering intraday, etc. data manually is boring, tedious, and for the birds; but we do what we gotta do. I'm pacing myself (to coincide with my NN's S-Fund exit time-frame window). I expect to have the "Deadline-NN" operational by or about the 20th of this month.

100% S by cob 6-3-14 (shudda been today)

Found a free source of intraday data going back to 9-9-2011. It required MUCH massaging and cleaning and it was for the Nasdaq Composite. Finally created a model. Tested the model. Unfortunately, the model was less accurate than my current end-of-day data models. Perhaps this means there is more noise/randomness on the intraday level, OR that more samples are needed OR that their is not enough of a correlation between the IXIC and DWCPF. Project scrapped for now.

For the record:

Over the next several days (weeks?), I will attempt to codify a variable Zig Zag indicator for use (after optimization) as a predictive model and then, realistically back-test/compare it to my current best model (born in the latter part of May and very successful in realistic back-testing, and, doing well since going online).

My current model also uses a non-lagging smoothing custom approach; however, this Zig Zag indicator, albeit much more difficult to program, seems to be even smoother--also with no lag (unlike moving averages, etc.).

The smoothest is a straight line of best fit since the markets began trading until today. The choppiest is tick data. The sweet spot lies somewhere in between, I believe. An optimal trading time frame; with an optimal resolution or scope.

ZigZag - ChartSchool - StockCharts.com

Originally Posted by userque

For the record:

Over the next several days (weeks?), I will attempt to codify a variable Zig Zag indicator for use (after optimization) as a predictive model and then, realistically back-test/compare it to my current best model (born in the latter part of May and very successful in realistic back-testing, and, doing well since going online).

My current model also uses a non-lagging smoothing custom approach; however, this Zig Zag indicator, albeit much more difficult to program, seems to be even smoother--also with no lag (unlike moving averages, etc.).

The smoothest is a straight line of best fit since the markets began trading until today. The choppiest is tick data. The sweet spot lies somewhere in between, I believe. An optimal trading time frame; with an optimal resolution or scope.

ZigZag - ChartSchool - StockCharts.com

Ha, codifying the zig zag was not the monster I'd read it would be. Finished that and finished back-testing it also. It didn't do well. On to the RSI.

The RSI is the only other indicator that I (currently) believe has a chance to beat my best model so far. It will be optimized based upon the number of days to use; trigger levels (for example, greater than 70%, but less than 90% for a short entry signal); and the same for an exit signal (less than 55%, exit to cash, for example. Or stay in until an opposite entry signal occurs. Or exit to cash after a variable, but set in advance number of days.)

Codifying the RSI is easy, but back-testing it will be the most time-consuming as there are many parameters with many possibilities each.

Originally Posted by userque

Ha, codifying the zig zag was not the monster I'd read it would be. Finished that and finished back-testing it also. It didn't do well. On to the RSI.

The RSI is the only other indicator that I (currently) believe has a chance to beat my best model so far. It will be optimized based upon the number of days to use; trigger levels (for example, greater than 70%, but less than 90% for a short entry signal); and the same for an exit signal (less than 55%, exit to cash, for example. Or stay in until an opposite entry signal occurs. Or exit to cash after a variable, but set in advance number of days.)

Codifying the RSI is easy, but back-testing it will be the most time-consuming as there are many parameters with many possibilities each.

Nope. RSI didn't fair well over a long time span.

Curiosities satisfied.

This past weekend, I made major progress by slightly increasing the 'resolution' of my pattern matching algorithms. Realistic back-testing using a simulated trading account spreadsheet including fees turned $1,000 into $2,496.44 from start of year until 6-16-14 by trading 3X W4500 (to estimate trading TNA and TZA. Actual TNA/TZA trades would likely yield greater gains.).

Increasing the resolution even further yielded worse results. Hopefully, I'm at a global best value and not just a local one.

I also had an epiphany regarding my TSP trading system. That, along with the above upgrade, made the TSP system up over 20% since start of year until 6-16-14. This doesn't include the likely F and G fund gains (and losses?) it would have encountered by flowing into the F and G funds when it predicts a downturn in the S fund. (It only includes [long, of course] trades into and out-of the S-Fund.)

For comparison to the Challenge, if I were using this latest system since the start of the Challenge, it would have turned the $10,000 into $16,183.95 so far.

I've thrown a lot of math at the problem of 'how best to trade the markets:' Artificial Neural Networks, Straight Statistics, Excel 2013 Solver using GRG and Evolutionary Algorithms, etc.

I can't think of any other (likely fruitful) trading concepts to test nor any other ways to test them.

I may start posting TNA/TZA trades here for a bit as an illustration and/or 'evidence of concept' as some can't/don't appreciate the legitimacy of proper back-testing. Other than that, and for now,

It is finished.