Each neural network has 15 inputs, 7 neurons on a single hidden layer and 1 output. 1) Ave no of trades/day/symbol It takes 2.2 trades per day per symbol 2) Ave hold time Hold time is under 1 hour 3) Calmar Ratio (Annualized Return / Max Drawdown) I am not familiar with this ratio. Will compute and get back to you 4) In vs out sample returns I will get to this when I post some of the equity curves, but quick answer is out of sample actually performs better. Considering both NQ and CL are completely out of sample. But even if you just look at 2009 until now on ES those results are better than 2000-2009 which was the only in-sample period
Thanks very much for the stats, Frost. So assuming you're not doing stuff like snipping connections during optimization, that's (15 * 7) + 7 = 112 weights per NN, and with 20 nets total that comes to 2240 tunable parameters (this doesn't include the bias of each neuron). That a lot of degrees of freedom. One thing to consider is that even though NQ and CL are different symbols than the one (ES) with which you tuned your NNs, you can't totally claim that they are purely out of sample, especially NQ which is very correlated to ES. Even though they are different symbols, you are using the same market period/regimes for your out of sample test. But if your bot is performing as well in the completely out of sample years as you say, then perhaps my points are moot. Congrats on your discovery! Look forward to seeing how things progress for you.
To remove all doubt in my mind before turning this on live, I decided to trade against a few more instruments to determine if the edge would persist: ZB: 6,281 trades, 68.23% profitable, +$183,985 per contract over 10 years.. Significantly lower performance than CL,ES, or NQ.. but still profitable ZN: 6,327 trades 72.06% profitable, +$121,240 per contract over 10 years. ZF:6,146 trades 70.47% profitable, +$76,168 per contract 6E: 6865 trades, 83.55% profitable, +$391,306 per contract. I am not very familiar with 6E does it have sufficient spreads/volume? It performs pretty well. Its a good sign that the strategy is profitable over all these instruments that I have never even tested against. All these results include commission and 1 tick of slippage. Are there any other instruments someone would like me to test against?
When backtesting your fills are simulated as of the current market price. So with ES for example, the last trade may have been 1225.25, but you do not know if that price was the bid or the ask. If you go to buy and the 1225.25 was at the ask, then you get filled at 1225.25 with your market order.. BUT if it was the bid, then you get filled at 1225.50 instead. Therefore you incur a tick of slippage. In order to trust the results I get from backtesting going forward, I have to add some slippage to each trade.
You could try TF, although its returns will correlate to ES. More volatility though. If you wanted to try some international indices, you could test the thicker ones, ESTX50, K200, and maybe DAX. ESTX50 does similar volume to ES, and K200 does about 500,000 contracts/day. ESTX has about 5x the daily range as ES. 6E has a nice daily range/margin ratio and does some decent volume (about 350,000 contracts/day last I checked). Certainly not as thick as ES though. Just curious: What brokerage are you planning on trading through? What's their intraday margin on ES? (I currently trade through IB via a custom API)
I need to look into it more and perhaps even do some tweaking, but I may replace NQ with 6E. Gives me slightly better diversification even though its a slight hit to the results.
Just a few questions from a discretionary trader... (1) What do you mean by 'flat at the close'? (2) Will you enter the trades manually (use the robot as the alerting system)?