كيك
How do I know if i Have discovered an edge?
Kayden Cook
Juan Smith
*hedge
Connor Brown
If you're using OHLC data, make sure you're not leaking future information. Remember that the close doesn't take place until... Well, the close of the trading day (or observational period in your data). You need to shift all your data down one row.
Jonathan Perry
Also I'm under the suspicion you're using an ensemble model (tree based?), looks like you have number of learners and some regularization term as hyperparams.
Keep in mind tree based ensembles CANNOT EXTRAPOLATE meaning they'll fail if presented with data they've never seen before. But this can be overcome by clever feature engineering.
To see if indeed you found alpha, you need to backtest it, mimicking stop losses on each trading period (day, hour, etc). Good luck user
Dylan Long
yeh of course im not doing that.
no its nothing fancy. Just brute force try every single param and build a graph. I already have the strategy. Just finding optimal parameters. The question i'm asking is, if my strategy is better than random would the curves look like mine do, or something different? I cba to do a pure random walk and build all the graphs and shit again multiple times.
Jacob Hernandez
>To see if indeed you found alpha, you need to backtest it, mimicking stop losses on each trading period (day, hour, etc). Good luck user
Ok I will build this graph now and let me know what you think? just a sec.... backtesting is done in javascript so its slow as fuck lol. cba to change it all to rust...
Elijah Williams
X axis is timeframe. (1 min candles - 30 min candles). Y axis is equity.
Even though the far right looks bad, It's still 1.012 equity. So 1.2% returns over a few months. No losses on any timeframe. Data was has been overfitted to the 5 min though but even without overfitting i suspect results will be similar but will look something like the graph in my next post
Leo Reyes
whoops got the pics the wrong way round. This is the one with overfitting. The one above is not.
Tyler Jones
see you tomorrow
Noah Jones
Nice.
Let me clarify my previous remarks, to backtest you need to simulate entry/exits. You can do this in a table. In Python, you'd pass your observations into your model.predict() and the result would be your entry, exit, hold decision (assuming that's how your model works; in, out). Now, if you're model says to buy, then you'd have to calculate entry price - exit price (the observation/row where model.predict = exit). During the time periods between enter and exit, if price fell