Unlocking Market Secrets: Why Volume Bars Crush Standard Timeframes 🚀
Hey everyone! 👋
I’ve recently started exploring an exciting new direction in trading: Volume Bars. 📊 Instead of relying on traditional time-based candles that tick away regardless of market activity, Volume Bars form a new bar only after a set volume of assets is traded. 💡 This concept was famously detailed by Marcos López de Prado, and the logic is brilliant: it cuts out market noise during dead hours and zooms in on institutional activity when high-volatility moves actually happen. 🚀
Looking at the initial results, training and validation performance on Volume Bars vastly outpaces anything I've seen on standard timeframes. 📈
02/25 on gpu1 | Epoch: 12 | TEST Acc: 0.5053 | TEST AUC: 0.4992 | TEST Loss: 0.8281
01/25 on gpu0 | Epoch: 07 | TEST Acc: 0.6056 | TEST AUC: 0.6491 | TEST Loss: 0.6810
04/25 on gpu1 | Epoch: 18 | TEST Acc: 0.6658 | TEST AUC: 0.7327 | TEST Loss: 0.6363
05/25 on gpu0 | Epoch: 20 | TEST Acc: 0.6939 | TEST AUC: 0.7126 | TEST Loss: 0.5996
03/25 on gpu0 | Epoch: 17 | TEST Acc: 0.5735 | TEST AUC: 0.5900 | TEST Loss: 1.0926
06/25 on gpu1 | Epoch: 20 | TEST Acc: 0.6110 | TEST AUC: 0.5811 | TEST Loss: 0.7518
08/25 on gpu1 | Epoch: 09 | TEST Acc: 0.6324 | TEST AUC: 0.6539 | TEST Loss: 0.6487
07/25 on gpu0 | Epoch: 06 | TEST Acc: 0.5816 | TEST AUC: 0.5859 | TEST Loss: 0.7404
10/25 on gpu1 | Epoch: 11 | TEST Acc: 0.5254 | TEST AUC: 0.5940 | TEST Loss: 0.7342
09/25 on gpu0 | Epoch: 22 | TEST Acc: 0.5147 | TEST AUC: 0.5110 | TEST Loss: 0.9248
12/25 on gpu1 | Epoch: 12 | TEST Acc: 0.6671 | TEST AUC: 0.6082 | TEST Loss: 0.6589
11/25 on gpu0 | Epoch: 24 | TEST Acc: 0.5067 | TEST AUC: 0.5770 | TEST Loss: 0.8748
14/25 on gpu1 | Epoch: 17 | TEST Acc: 0.6524 | TEST AUC: 0.6489 | TEST Loss: 0.6689
13/25 on gpu0 | Epoch: 23 | TEST Acc: 0.6457 | TEST AUC: 0.6759 | TEST Loss: 0.6529
15/25 on gpu0 | Epoch: 07 | TEST Acc: 0.5441 | TEST AUC: 0.6295 | TEST Loss: 0.7677
16/25 on gpu1 | Epoch: 07 | TEST Acc: 0.5481 | TEST AUC: 0.6256 | TEST Loss: 0.8671
18/25 on gpu1 | Epoch: 16 | TEST Acc: 0.6417 | TEST AUC: 0.7187 | TEST Loss: 0.6319
19/25 on gpu0 | Epoch: 27 | TEST Acc: 0.6471 | TEST AUC: 0.6415 | TEST Loss: 0.8184
17/25 on gpu0 | Epoch: 39 | TEST Acc: 0.6310 | TEST AUC: 0.6684 | TEST Loss: 0.8783
20/25 on gpu1 | Epoch: 09 | TEST Acc: 0.6511 | TEST AUC: 0.6779 | TEST Loss: 0.6568
21/25 on gpu0 | Epoch: 21 | TEST Acc: 0.6096 | TEST AUC: 0.6774 | TEST Loss: 0.6528
22/25 on gpu1 | Epoch: 21 | TEST Acc: 0.4693 | TEST AUC: 0.6018 | TEST Loss: 0.9780
23/25 on gpu0 | Epoch: 11 | TEST Acc: 0.5575 | TEST AUC: 0.5698 | TEST Loss: 0.7762
24/25 on gpu1 | Epoch: 07 | TEST Acc: 0.5695 | TEST AUC: 0.6324 | TEST Loss: 0.6885
25/25 on gpu0 | Epoch: 06 | TEST Acc: 0.6230 | TEST AUC: 0.6470 | TEST Loss: 0.6962
Analyzing the metrics across several training approaches, the AUC scores—which measure the model’s ability to separate signals from noise—are consistently hitting impressive territory, far surpassing the typical 0.53–0.55 range that is usually considered solid for Forex ML models. 🎯
I’m confident this is shaping up to be my strongest strategy yet, which is why I’m officially naming it Volume Reaper! 🦾
My next step is to attach a Reinforcement Learning (RL) model on top of this framework to double-check signals and optimize directly for PnL quality. 🧠 Stay tuned for more updates! 🔥
