Bug in MACD Signal: repainting the last two candles - lookahead bias in backtest

[hotston]

I noticed that MACD Signal always repaints for the last two candles, forcing a lookahead bias in backtest. In live/dry run it would always show up as 0 or an extremely small value, making it useless.

Steps to reproduce:

1. Use freqtrade or similar software
2. Use a strategy with MACD Signal, such as: https://github.com/davidzr/freqtrade-strategies/blob/main/strategies/TheForce/TheForce.py
3. Log the values, observe the following. MACD Signal is 0 for the past two candles.

2025-03-16 06:41:06,339 - TheForce5m - INFO - [DEBUG] NEAR/USDT:USDT - 2025-03-16 04:30:00+00:00 | Close: 2.64 | MACD: -0.0038208040952749656 | MACD Signal: -6.701088123062249e-06 | FastK: 62.06896551724162 | FastD: 67.56045052940304
2025-03-16 06:41:06,341 - TheForce5m - INFO - [DEBUG] NEAR/USDT:USDT - 2025-03-16 05:00:00+00:00 | Close: 2.655 | MACD: -0.0021882205918517528 | MACD Signal: 0.0003464176237795158 | FastK: 85.41666666666643 | FastD: 81.00651439338785
2025-03-16 06:41:06,343 - TheForce5m - INFO - [DEBUG] NEAR/USDT:USDT - 2025-03-16 05:30:00+00:00 | Close: 2.666 | MACD: -6.701088123062249e-06 | MACD Signal: 0.0 | FastK: 97.22222222222254 | FastD: 94.32473509437078
2025-03-16 06:41:06,345 - TheForce5m - INFO - [DEBUG] NEAR/USDT:USDT - 2025-03-16 06:00:00+00:00 | Close: 2.649 | MACD: 0.0003464176237795158 | MACD Signal: 0.0 | FastK: 50.00000000000062 | FastD: 62.961995384913024

Expected behaviour:

MACD Signal should be based on the historical data and should never be forward looking, creating a lookahead bias.

[mario4tier]

There is a pending change at some point this year for TA-Lib to fix period=1.

You will find more details in (#48).

Please give it a shot with period=2 in the context of your application and let us know.

[hotston]

TA-Lib

yeah, period=2 works. Very strangy bug

[mario4tier]

Thanks for verifying.

The bug is well understood at this point. The reasons for the long time taken to fix this is:

(1) It is not typical to use a period of "1" for a signal line... usually the algos want the signal line to be "slower". So most users are not affected by this edge case.

(2) I am very very careful+nervous about touching "fundamental" functions such as MACD because these are time tested (many businesses are running just fine for years with the current code).

... but I recognize it is overdue to make the change because enough people have stumble on this. Very likely this will be fix this year.