ANN based Quantitative Trading Strategy

Pine ScriptMachine learning Trend following Volatility

Author: ChaoZhang
Complexity: Starter · 83 lines
Parameters: 2
Backtest bar: 1h · intraday
Tested on: BTC/USDT · futures

Jump to code View in vault Original on FMZ

Overview

This strategy uses an artificial neural network (ANN) to predict future price changes and generates trading signals based on the predictions. It belongs to the trend following strategy. The advantage is that it can identify complex nonlinear trends and is suitable for medium-to-long term trading. However, it also has the risk of overfitting to backtest data and underperforming in live trading.

Strategy Logic

The strategy uses an ANN model to predict the percentage change of the next trading day.

The input layer has only one node, which is the percentage change of the previous day.

The hidden layer has 2 layers, the first with 5 nodes, and the second with 33 nodes. Both use hyperbolic tangent (tanh) as activation function.

The output layer has one node, going through a linear activation function to generate the final prediction.

If the prediction is greater than the threshold parameter, a long signal is generated. If less than the negative of threshold, a short signal is generated.

Advantages

ANN can model complex nonlinear relationships in data
Only requires data from previous day as input
Can identify longer timeframes trends
Multiple hidden layers increase modeling capability
Activation functions and parameters optimized for good performance

Risks

Overfitting risk - live performance may differ from backtest
Requires longer historical data for training
Parameters and structure need optimization, results may vary
Only predicts next day, unable to determine long term trends
Performance may degrade during ranging markets

Improvement Directions

Add more input variables like volume etc
Try different ANN architectures and activation functions
Optimize network parameters for better fit
Increase sample size of training data to reduce overfitting
Predict across multiple time horizons to better identify trends
Ensemble with other models
Use volatility measures etc for better risk control

Conclusion

This ANN-based strategy can identify complex nonlinear trends and is suitable for medium-to-long term trading. However, the black-box nature of ANN models also poses significant challenges for live trading. We need to optimize across input features, model architecture, parameter tuning, ensemble learning etc. while combining with traditional technical analysis for robust real-world performance. AI strategies still need to blend with conventional techniques to maximize performance.

Parameters

Name	Default	What it controls
v_input_1	false	Threshold
v_input_2	1D	Timeframe

Source Pine Script

/*backtest
start: 2023-10-14 00:00:00
end: 2023-11-13 00:00:00
period: 1h
basePeriod: 15m
exchanges: [{"eid":"Futures_Binance","currency":"BTC_USDT"}]
*/

//@version=2
strategy("ANN Strategy v2")

threshold = input(title="Threshold", type=float, defval=0.0000, step=0.0001)
timeframe = input(title="Timeframe",  defval='1D' )

getDiff() =>
    yesterday=request.security(syminfo.tickerid, timeframe, ohlc4[1])
    today=ohlc4
    delta=today-yesterday
    percentage=delta/yesterday

PineActivationFunctionLinear(v) => v
PineActivationFunctionTanh(v) => (exp(v) - exp(-v))/(exp(v) + exp(-v))


l0_0 = PineActivationFunctionLinear(getDiff())

l1_0 = PineActivationFunctionTanh(l0_0*0.8446488687)
l1_1 = PineActivationFunctionTanh(l0_0*-0.5674069006)
l1_2 = PineActivationFunctionTanh(l0_0*0.8676766445)
l1_3 = PineActivationFunctionTanh(l0_0*0.5200611473)
l1_4 = PineActivationFunctionTanh(l0_0*-0.2215499554)

l2_0 = PineActivationFunctionTanh(l1_0*0.3341657935 + l1_1*-2.0060003664 + l1_2*0.8606354375 + l1_3*0.9184846912 + l1_4*-0.8531172267)
l2_1 = PineActivationFunctionTanh(l1_0*-0.0394076437 + l1_1*-0.4720374911 + l1_2*0.2900968524 + l1_3*1.0653326022 + l1_4*0.3000188806)
l2_2 = PineActivationFunctionTanh(l1_0*-0.559307785 + l1_1*-0.9353655177 + l1_2*1.2133832962 + l1_3*0.1952686024 + l1_4*0.8552068166)
l2_3 = PineActivationFunctionTanh(l1_0*-0.4293220754 + l1_1*0.8484259409 + l1_2*-0.7154087313 + l1_3*0.1102971055 + l1_4*0.2279392724)
l2_4 = PineActivationFunctionTanh(l1_0*0.9111779155 + l1_1*0.2801691115 + l1_2*0.0039982713 + l1_3*-0.5648257117 + l1_4*0.3281705155)
l2_5 = PineActivationFunctionTanh(l1_0*-0.2963954503 + l1_1*0.4046532178 + l1_2*0.2460580977 + l1_3*0.6608675819 + l1_4*-0.8732022547)
l2_6 = PineActivationFunctionTanh(l1_0*0.8810811932 + l1_1*0.6903706878 + l1_2*-0.5953059103 + l1_3*-0.3084040686 + l1_4*-0.4038498853)
l2_7 = PineActivationFunctionTanh(l1_0*-0.5687101164 + l1_1*0.2736758588 + l1_2*-0.2217360382 + l1_3*0.8742950972 + l1_4*0.2997583987)
l2_8 = PineActivationFunctionTanh(l1_0*0.0708459913 + l1_1*0.8221730616 + l1_2*-0.7213265567 + l1_3*-0.3810462836 + l1_4*0.0503867753)
l2_9 = PineActivationFunctionTanh(l1_0*0.4880140595 + l1_1*0.9466627196 + l1_2*1.0163097961 + l1_3*-0.9500386514 + l1_4*-0.6341709382)
l2_10 = PineActivationFunctionTanh(l1_0*1.3402207103 + l1_1*0.0013395288 + l1_2*3.4813009133 + l1_3*-0.8636814677 + l1_4*41.3171047132)
l2_11 = PineActivationFunctionTanh(l1_0*1.2388217292 + l1_1*-0.6520886912 + l1_2*0.3508321737 + l1_3*0.6640560714 + l1_4*1.5936220597)
l2_12 = PineActivationFunctionTanh(l1_0*-0.1800525171 + l1_1*-0.2620989752 + l1_2*0.056675277 + l1_3*-0.5045395315 + l1_4*0.2732553554)
l2_13 = PineActivationFunctionTanh(l1_0*-0.7776331454 + l1_1*0.1895231137 + l1_2*0.5384918862 + l1_3*0.093711904 + l1_4*-0.3725627758)
l2_14 = PineActivationFunctionTanh(l1_0*-0.3181583022 + l1_1*0.2467979854 + l1_2*0.4341718676 + l1_3*-0.7277619935 + l1_4*0.1799381758)
l2_15 = PineActivationFunctionTanh(l1_0*-0.5558227731 + l1_1*0.3666152536 + l1_2*0.1538243225 + l1_3*-0.8915928174 + l1_4*-0.7659355684)
l2_16 = PineActivationFunctionTanh(l1_0*0.6111516061 + l1_1*-0.5459495224 + l1_2*-0.5724238425 + l1_3*-0.8553500765 + l1_4*-0.8696190472)
l2_17 = PineActivationFunctionTanh(l1_0*0.6843667454 + l1_1*0.408652181 + l1_2*-0.8830470112 + l1_3*-0.8602324935 + l1_4*0.1135462621)
l2_18 = PineActivationFunctionTanh(l1_0*-0.1569048216 + l1_1*-1.4643247888 + l1_2*0.5557152813 + l1_3*1.0482791924 + l1_4*1.4523116833)
l2_19 = PineActivationFunctionTanh(l1_0*0.5207514017 + l1_1*-0.2734444192 + l1_2*-0.3328660936 + l1_3*-0.7941515963 + l1_4*-0.3536051491)
l2_20 = PineActivationFunctionTanh(l1_0*-0.4097807954 + l1_1*0.3198619826 + l1_2*0.461681627 + l1_3*-0.1135575498 + l1_4*0.7103339851)
l2_21 = PineActivationFunctionTanh(l1_0*-0.8725014237 + l1_1*-1.0312091401 + l1_2*0.2267643037 + l1_3*-0.6814258121 + l1_4*0.7524828703)
l2_22 = PineActivationFunctionTanh(l1_0*-0.3986855003 + l1_1*0.4962556631 + l1_2*-0.7330224516 + l1_3*0.7355772164 + l1_4*0.3180141739)
l2_23 = PineActivationFunctionTanh(l1_0*-1.083080442 + l1_1*1.8752543187 + l1_2*0.3623326265 + l1_3*-0.348145191 + l1_4*0.1977935038)
l2_24 = PineActivationFunctionTanh(l1_0*-0.0291290625 + l1_1*0.0612906199 + l1_2*0.1219696687 + l1_3*-1.0273685429 + l1_4*0.0872219768)
l2_25 = PineActivationFunctionTanh(l1_0*0.931791094 + l1_1*-0.313753684 + l1_2*-0.3028724837 + l1_3*0.7387076712 + l1_4*0.3806140391)
l2_26 = PineActivationFunctionTanh(l1_0*0.2630619402 + l1_1*-1.9827996702 + l1_2*-0.7741413496 + l1_3*0.1262957444 + l1_4*0.2248777886)
l2_27 = PineActivationFunctionTanh(l1_0*-0.2666322362 + l1_1*-1.124654664 + l1_2*0.7288282621 + l1_3*-0.1384289204 + l1_4*0.2395966188)
l2_28 = PineActivationFunctionTanh(l1_0*0.6611845175 + l1_1*0.0466048937 + l1_2*-0.1980999993 + l1_3*0.8152350927 + l1_4*0.0032723211)
l2_29 = PineActivationFunctionTanh(l1_0*-0.3150344751 + l1_1*0.1391754608 + l1_2*0.5462816249 + l1_3*-0.7952302364 + l1_4*-0.7520712378)
l2_30 = PineActivationFunctionTanh(l1_0*-0.0576916066 + l1_1*0.3678415302 + l1_2*0.6802537378 + l1_3*1.1437036331 + l1_4*-0.8637405666)
l2_31 = PineActivationFunctionTanh(l1_0*0.7016273068 + l1_1*0.3978601709 + l1_2*0.3157049654 + l1_3*-0.2528455662 + l1_4*-0.8614146703)
l2_32 = PineActivationFunctionTanh(l1_0*1.1741126834 + l1_1*-1.4046408959 + l1_2*1.2914477803 + l1_3*0.9904052964 + l1_4*-0.6980155826)

l3_0 = PineActivationFunctionTanh(l2_0*-0.1366382003 + l2_1*0.8161960822 + l2_2*-0.9458773183 + l2_3*0.4692969576 + l2_4*0.0126710629 + l2_5*-0.0403001012 + l2_6*-0.0116244898 + l2_7*-0.4874816289 + l2_8*-0.6392241448 + l2_9*-0.410338398 + l2_10*-0.1181027081 + l2_11*0.1075562037 + l2_12*-0.5948728252 + l2_13*0.5593677345 + l2_14*-0.3642935247 + l2_15*-0.2867603217 + l2_16*0.142250271 + l2_17*-0.0535698019 + l2_18*-0.034007685 + l2_19*-0.3594532426 + l2_20*0.2551095195 + l2_21*0.4214344983 + l2_22*0.8941621336 + l2_23*0.6283377368 + l2_24*-0.7138020667 + l2_25*-0.1426738249 + l2_26*0.172671223 + l2_27*0.0714824385 + l2_28*-0.3268182144 + l2_29*-0.0078989755 + l2_30*-0.2032828145 + l2_31*-0.0260631534 + l2_32*0.4918037012)

buying = l3_0 > 0 ? true : l3_0 < -0 ? false : buying[1]

hline(0, title="base line")
//bgcolor(l3_0 > 0.0014 ? green : l3_0 < -0.0014 ? red : gray, transp=20)
bgcolor(buying ? green : red, transp=20)
plot(l3_0, color=silver, style=area, transp=75)
plot(l3_0, color=aqua, title="prediction")

longCondition = buying
if (longCondition)
    strategy.entry("Long", strategy.long)

shortCondition = buying != true
if (shortCondition)
    strategy.entry("Short", strategy.short)

Published by ChaoZhang on the FMZ strategy square · last modified 2023-11-14. Collected in the Quant Trading Vault. Unvalidated research: re-test with your own data and costs before trusting it.

Build this with AgenKit

A spec is not a trading system. Paste this into your coding agent and AgenKit runs a specialist team — architecture, test-first build, risk gates, review — with paper trading on by default.

in Claude Code, Cursor, Codex & 4 more

/agenkit build a production-grade trading bot with a backtest engine, risk controls and paper trading by default from strategies/ANN-based-Quantitative-Trading-Strategy.md

npx agenkit install engineering-kit in your project
Paste the command into your agent and approve each phase
Get tests, backtest, risk controls and a paper-trading switch

Get AgenKit How /agenkit works

Understand it first

Before you run this, read the lesson behind the idea and how to backtest it honestly. Specs like this one are the survivors of an unknown number of trials; deflate their results accordingly.