⚡️ New trend detection hilbert transform

[raftersvk]

No description provided.

[raftersvk]

Hi,

following the first idea of using SAR to improve trend detection I decided to give a go at another indicator :
Hilbert Transform Trend vs Cycle
https://www.tradingpedia.com/videos/trend-vs-cycle-mode-understanding-hilbert-transform/

below is my first test of 800 epoch without ROI and trailing in order to really compare the buy sell impact.

here is a comparison of the best results : New Trend on the left side / Vanilla (ADX) on the right side

• first we can not see a big difference on total profit
• but average duration seems better for the same kind of profit
  

this is confirmed when looking at epoch with duration less than a day :

detailed files / log here :
MGM_v012_HTTrend_comparison.zip

[raftersvk]

there is probably one improvment on this trend detection by avoiding setting the trend which last only one candle as seen below :

if anyone can help me do that ;)

even further : could the trend "duration" be an hyperotable parameter ?

[Rikj000]

there is probably one improvement on this trend detection by avoiding setting the trend which last only one candle as seen below
if anyone can help me do that ;)

I'm thinking perhaps we could do this particular improvement at a later time! When I implement: #44
Once the features are in place to make multiple candle usage easy, then we could use our core trend indicator at larger candles, which should yield robuster / slower moving trends. I think that should reduce this issue from occuring 🙂

[Rikj000]

Good spot! 😄 But this will require new tests then! 👀

[raftersvk]

indeed, so here are the latest results of the night ;)

total profit comparison : SharpeLoss newTrend left / SharpeLoss Vanilla right

=> total profit is not so good as first test, but not that far from vanilla either

epochs lasting less than a day :

=> still better result than vanilla, but not as strong as first test

epoch during between 24h and 48h :

=> tendency for shorter duration and better result is confirmed here with newTrend

files / log here :
MGM_v012_noRoiNoTrail_Sharpe_HTTREND_v2.zip

[raftersvk]

same analysis with SortinoLoss instead of Sharpe vs Vanilla(Sharpe)

total profit

=> still competitive new trend

epochs < 24h

=> new trend better and also better than SharpeLoss (not sure what it means or if it only depends of the timerange)

epochs >= 24h < 48h

=> still better than vanilla but not as good as Sharpe this time

files / log :
MGM_v012_noRoiNoTrail_Sortino_HTTREND_v2.zip

[raftersvk]

to complete the picture here is a plot of the underlying trend detection :

• subplot 1 : HT Trend
• subplot 2 : resulting MGM Trend (-1 = downwards / 0 = sideways / 1 = upwards)
• subplot 3 : plusDI / minusDI (to get the direction of the trend)

[raftersvk]

continuing my research, this is a new plot with the following config :
"trading_during_trends": {
"buy_trades_when_downwards": false,
"buy_trades_when_sideways": false,
"buy_trades_when_upwards": true,
"sell_trades_when_downwards": true,
"sell_trades_when_sideways": true,
"sell_trades_when_upwards": true
}

on a big picture it looks like we can avoid the big falls, but also miss some of the rebound in the downtrend.

freqtrade-plot-BTC_USDT-1h.zip

let's see what we get with different setups ;)

[Rikj000]

Merged with: #79