Oooh, like the not tested against a benchmark time series. That would definitely make good garbage. Ok, was afraid I needed to check deeper here.
I have started training a new class of daytrading strategies. My previous strategies are largely a "black box". Very little guidance was given to the system as far what type of strategy it could produce. However, this new class of strategies I am training around a "base concept" Base Strategy: - The strategy analyzes the first 15-20 minutes of each session - The strategy then decides to go long/short or do nothing for the day - The strategy exits once it gets a 1 min bar close that is > 1% or < -1% from the entry - The strategy exits roughly 20 minutes before market close if it failed to move +-/1% So for each symbol being traded, it's just 1 trade per day. I am training using the same inputs as my other strategies (which as discussed I will not disclose). Because of how the base strategy is bounded, I can actually train about 1 strategy per day. For now, each day I will be training for a new symbol. The first symbol I trained on was AAPL. I completed training on Monday night and ran the strategy live yesterday. It resulted in a win and a gain of roughly $200 (Putting 20k into each position right now). Last night I completed training on MSFT and ran it live today. Today AAPL won again for another $200 gain and MSFT lost for a -$20. I am actually posting these trades live as they happen to my twitter account. Since its just 1 position per instrument per day, it's really easy to follow what the strategy is doing. I am hoping to complete training of CAT tonight and put it live tomorrow. I have attached a PL/Drawdown chart for the AAPL strategy. It's essentially averaging 50% annual returns against an 8% max drawdown with most years sub 5%. These are very different strategies compared to what I normally do. I am curious to see how these work out as I am essentially forward testing live.
May I suggest BTU as a contender, it appears to me to have good moves. I have no interest or trades in BTU.
Why not just exit with a Market on Close order? Are you doing stocks as well as Futures in order to diversify even more? Or you think that you can get a better return? how do you calculate expectancy percentage? Is it relative to your margin usage? What surprises me is that your models are stock specific (they probably have market features though). Anyway whatever works for you!
A noob question: If the trading strategies are auto-generated, is that being produced for each trade or for the entire trading period, be it a day or a month (namely, is it static or dynamic)? That said, do you know or does anyone know a way to have the ML auto-generate an algo on a "per candle/bar" basis and have the data backtested (within a reasonable timeframe like the previous 30 days) in the interim?
Interesting. AAPL,MSFT, and CAT - None of those new strategies triggered today. Perhaps that says something about the anticipated market action today. Working on adding AXP tonight.
AXP and MSFT were the two that fired today. Hit profit target on both. Made $209 on MSFT and $106 on AXP. In my rush to get AXP turned on, I accidentally had it set to only do $10k per position instead of $20k per position like the others. For tonight I am training TGT. Haven't decided what I am training tomorrow yet. Overall this week, this new strategy concept had 5 trades, 4 winners and 1 loser for a gain of: $711