11188 lượt xem
1170
11188 16
This is an implementation of an Artificial Neural Network (ANN) in pine. I made this as part of a bigger project and should be considered more as a proof of concept than a fully working indicator.

It was trained by a different program, using 3 years of bitcoin history. It's a 4 layer ANN that takes the percentual difference of the last few days as input. It was randomly generated (that's why the input is a bit funny :) ) and translated to pine.

It seems to be much better at trends and pump/dumps, because it keeps switching during sideways, so perhaps with a sideways indicator this could actually prove useful. In the chart above, I drew how you would trade with this indicator. Every red square is a short, every green square is a long.

The aqua plot line is the actual prediction. If this is above 0, the background is drawn green and the indicator is bullish . Otherwise, the background is drawn red and the indicator is bearish .

Any and all question can be asked by PM'ing me or emailing me at masterflappie@gmail.com
Gỡ bỏ khỏi Script Ưa thích Thêm vào Script Ưa thích
study("ANN")

getDiff(offset) =>
    yesterday=close[offset+1]
    today=close[offset]
    delta=today-yesterday
    percentage=delta/yesterday
 
PineActivationFunctionLinear(v) => v
 
PineActivationFunctionTanh(v) => (exp(v) - exp(-v))/(exp(v) + exp(-v))

l0_0 = PineActivationFunctionLinear(getDiff(0))
l0_1 = PineActivationFunctionLinear(getDiff(0))
l0_2 = PineActivationFunctionLinear(getDiff(0))
l0_3 = PineActivationFunctionLinear(getDiff(0))
l0_4 = PineActivationFunctionLinear(getDiff(0))
l0_5 = PineActivationFunctionLinear(getDiff(0))
l0_6 = PineActivationFunctionLinear(getDiff(0))
l0_7 = PineActivationFunctionLinear(getDiff(0))
l0_8 = PineActivationFunctionLinear(getDiff(0))
l0_9 = PineActivationFunctionLinear(getDiff(0))
l0_10 = PineActivationFunctionLinear(getDiff(0))
l0_11 = PineActivationFunctionLinear(getDiff(0))
l0_12 = PineActivationFunctionLinear(getDiff(0))
l0_13 = PineActivationFunctionLinear(getDiff(0))
l0_14 = PineActivationFunctionLinear(getDiff(0))
 
l1_0 = PineActivationFunctionTanh(l0_0*5.040340774 + l0_1*-1.3025994088 + l0_2*19.4225543981 + l0_3*1.1796960423 + l0_4*2.4299395823 + l0_5*3.159003445 + l0_6*4.6844527551 + l0_7*-6.1079267196 + l0_8*-2.4952869198 + l0_9*-4.0966081154 + l0_10*-2.2432843111 + l0_11*-0.6105764807 + l0_12*-0.0775684605 + l0_13*-0.7984753138 + l0_14*3.4495907342)
l1_1 = PineActivationFunctionTanh(l0_0*5.9559031982 + l0_1*-3.1781960056 + l0_2*-1.6337491061 + l0_3*-4.3623166512 + l0_4*0.9061990402 + l0_5*-0.731285093 + l0_6*-6.2500232251 + l0_7*0.1356087758 + l0_8*-0.8570572885 + l0_9*-4.0161353298 + l0_10*1.5095552083 + l0_11*1.324789197 + l0_12*-0.1011973878 + l0_13*-2.3642090162 + l0_14*-0.7160862442)
l1_2 = PineActivationFunctionTanh(l0_0*4.4350881378 + l0_1*-2.8956461034 + l0_2*1.4199762607 + l0_3*-0.6436844261 + l0_4*1.1124274281 + l0_5*-4.0976954985 + l0_6*2.9317456342 + l0_7*0.0798318393 + l0_8*-5.5718144311 + l0_9*-0.6623352208 +l0_10*3.2405203222 + l0_11*-10.6253384513 + l0_12*4.7132919253 + l0_13*-5.7378151597 + l0_14*0.3164836695)
l1_3 = PineActivationFunctionTanh(l0_0*-6.1194605467 + l0_1*7.7935605604 + l0_2*-0.7587522153 + l0_3*9.8382495905 + l0_4*0.3274314734 + l0_5*1.8424796541 + l0_6*-1.2256355427 + l0_7*-1.5968600758 + l0_8*1.9937700922 + l0_9*5.0417809111 + l0_10*-1.9369944654 + l0_11*6.1013201778 + l0_12*1.5832910747 + l0_13*-2.148403244 + l0_14*1.5449437366)
l1_4 = PineActivationFunctionTanh(l0_0*3.5700040028 + l0_1*-4.4755892733 + l0_2*0.1526702072 + l0_3*-0.3553664401 + l0_4*-2.3777962662 + l0_5*-1.8098849587 + l0_6*-3.5198449134 + l0_7*-0.4369370497 + l0_8*2.3350169623 + l0_9*1.9328960346 + l0_10*1.1824141812 + l0_11*3.0565148049 + l0_12*-9.3253401534 + l0_13*1.6778555498 + l0_14*-3.045794332)
l1_5 = PineActivationFunctionTanh(l0_0*3.6784907623 + l0_1*1.1623683715 + l0_2*7.1366362145 + l0_3*-5.6756546585 + l0_4*12.7019884334 + l0_5*-1.2347823331 + l0_6*2.3656619827 + l0_7*-8.7191778213 + l0_8*-13.8089238753 + l0_9*5.4335943836 + l0_10*-8.1441181338 + l0_11*-10.5688113287 + l0_12*6.3964140758 + l0_13*-8.9714236223 + l0_14*-34.0255456929)
l1_6 = PineActivationFunctionTanh(l0_0*-0.4344517548 + l0_1*-3.8262167437 + l0_2*-0.2051098003 + l0_3*0.6844201221 + l0_4*1.1615893422 + l0_5*-0.404465314 + l0_6*-0.1465747632 + l0_7*-0.006282458 + l0_8*0.1585655487 + l0_9*1.1994484991 + l0_10*-0.9879081404 + l0_11*-0.3564970612 + l0_12*1.5814717823 + l0_13*-0.9614804676 + l0_14*0.9204822346)
l1_7 = PineActivationFunctionTanh(l0_0*-4.2700957175 + l0_1*9.4328591157 + l0_2*-4.3045548 + l0_3*5.0616868842 + l0_4*3.3388781058 + l0_5*-2.1885073225 + l0_6*-6.506301518 + l0_7*3.8429000108 + l0_8*-1.6872237349 + l0_9*2.4107095799 + l0_10*-3.0873985314 + l0_11*-2.8358325447 + l0_12*2.4044366491 + l0_13*0.636779082 + l0_14*-13.2173215035)
l1_8 = PineActivationFunctionTanh(l0_0*-8.3224697492 + l0_1*-9.4825530183 + l0_2*3.5294389835 + l0_3*0.1538618049 + l0_4*-13.5388631898 + l0_5*-0.1187936017 + l0_6*-8.4582741139 + l0_7*5.1566299292 + l0_8*10.345519938 + l0_9*2.9211759333 + l0_10*-5.0471804233 + l0_11*4.9255989983 + l0_12*-9.9626142544 + l0_13*23.0043143258 + l0_14*20.9391809343)
l1_9 = PineActivationFunctionTanh(l0_0*-0.9120518654 + l0_1*0.4991807488 + l0_2*-1.877244586 + l0_3*3.1416466525 + l0_4*1.063709676 + l0_5*0.5210126835 + l0_6*-4.9755780108 + l0_7*2.0336532347 + l0_8*-1.1793121093 + l0_9*-0.730664855 + l0_10*-2.3515987428 + l0_11*-0.1916546514 + l0_12*-2.2530340504 + l0_13*-0.2331829119 + l0_14*0.7216218149)
l1_10 = PineActivationFunctionTanh(l0_0*-5.2139618683 + l0_1*1.0663790028 + l0_2*1.8340834959 + l0_3*1.6248173447 + l0_4*-0.7663740145 + l0_5*0.1062788171 + l0_6*2.5288021501 + l0_7*-3.4066549066 + l0_8*-4.9497988755 + l0_9*-2.3060668143 + l0_10*-1.3962486274 + l0_11*0.6185583427 + l0_12*0.2625299576 + l0_13*2.0270246444 + l0_14*0.6372015811)
l1_11 = PineActivationFunctionTanh(l0_0*0.2020072665 + l0_1*0.3885852709 + l0_2*-0.1830248843 + l0_3*-1.2408598444 + l0_4*-0.6365798088 + l0_5*1.8736534268 + l0_6*0.656206442 + l0_7*-0.2987482678 + l0_8*-0.2017485963 + l0_9*-1.0604095303 + l0_10*0.239793356 + l0_11*-0.3614172938 + l0_12*0.2614678044 + l0_13*1.0083551762 + l0_14*-0.5473833797)
l1_12 = PineActivationFunctionTanh(l0_0*-0.4367517149 + l0_1*-10.0601304934 + l0_2*1.9240604838 + l0_3*-1.3192184047 + l0_4*-0.4564760159 + l0_5*-0.2965270368 + l0_6*-1.1407423613 + l0_7*2.0949647291 + l0_8*-5.8212599297 + l0_9*-1.3393321939 + l0_10*7.6624548265 + l0_11*1.1309391851 + l0_12*-0.141798054 + l0_13*5.1416736187 + l0_14*-1.8142503125)
l1_13 = PineActivationFunctionTanh(l0_0*1.103948336 + l0_1*-1.4592033032 + l0_2*0.6146278432 + l0_3*0.5040966421 + l0_4*-2.4276090772 + l0_5*-0.0432902426 + l0_6*-0.0044259999 + l0_7*-0.5961347308 + l0_8*0.3821026107 + l0_9*0.6169102373 +l0_10*-0.1469847611 + l0_11*-0.0717167683 + l0_12*-0.0352403695 + l0_13*1.2481310788 + l0_14*0.1339628411)
l1_14 = PineActivationFunctionTanh(l0_0*-9.8049980534 + l0_1*13.5481068519 + l0_2*-17.1362809025 + l0_3*0.7142100864 + l0_4*4.4759163422 + l0_5*4.5716161777 + l0_6*1.4290884628 + l0_7*8.3952862712 + l0_8*-7.1613700432 + l0_9*-3.3249489518+ l0_10*-0.7789587912 + l0_11*-1.7987628873 + l0_12*13.364752545 + l0_13*5.3947219678 + l0_14*12.5267547127)
l1_15 = PineActivationFunctionTanh(l0_0*0.9869461803 + l0_1*1.9473351905 + l0_2*2.032925759 + l0_3*7.4092080633 + l0_4*-1.9257741399 + l0_5*1.8153585328 + l0_6*1.1427866392 + l0_7*-0.3723167449 + l0_8*5.0009927384 + l0_9*-0.2275103411 + l0_10*2.8823012914 + l0_11*-3.0633141934 + l0_12*-2.785334815 + l0_13*2.727981E-4 + l0_14*-0.1253009512)
l1_16 = PineActivationFunctionTanh(l0_0*4.9418118585 + l0_1*-2.7538199876 + l0_2*-16.9887588104 + l0_3*8.8734475297 + l0_4*-16.3022734814 + l0_5*-4.562496601 + l0_6*-1.2944373699 + l0_7*-9.6022946986 + l0_8*-1.018393866 + l0_9*-11.4094515429 + l0_10*24.8483091382 + l0_11*-3.0031522277 + l0_12*0.1513114555 + l0_13*-6.7170487021 + l0_14*-14.7759227576)
l1_17 = PineActivationFunctionTanh(l0_0*5.5931454656 + l0_1*2.22272078 + l0_2*2.603416897 + l0_3*1.2661196599 + l0_4*-2.842826446 + l0_5*-7.9386099121 + l0_6*2.8278849111 + l0_7*-1.2289445238 + l0_8*4.571484248 + l0_9*0.9447425595 + l0_10*4.2890688351 + l0_11*-3.3228258483 + l0_12*4.8866215526 + l0_13*1.0693412194 + l0_14*-1.963203112)
l1_18 = PineActivationFunctionTanh(l0_0*0.2705520264 + l0_1*0.4002328199 + l0_2*0.1592515845 + l0_3*0.371893552 + l0_4*-1.6639467871 + l0_5*2.2887318884 + l0_6*-0.148633664 + l0_7*-0.6517792263 + l0_8*-0.0993032992 + l0_9*-0.964940376 + l0_10*0.1286342935 + l0_11*0.4869943595 + l0_12*1.4498648166 + l0_13*-0.3257333384 + l0_14*-1.3496419812)
l1_19 = PineActivationFunctionTanh(l0_0*-1.3223200798 + l0_1*-2.2505204324 + l0_2*0.8142804525 + l0_3*-0.848348177 + l0_4*0.7208860589 + l0_5*1.2033423756 + l0_6*-0.1403005786 + l0_7*0.2995941644 + l0_8*-1.1440473062 + l0_9*1.067752916 + l0_10*-1.2990534679 + l0_11*1.2588583869 + l0_12*0.7670409455 + l0_13*2.7895972983 + l0_14*-0.5376152512)
l1_20 = PineActivationFunctionTanh(l0_0*0.7382351572 + l0_1*-0.8778865631 + l0_2*1.0950766363 + l0_3*0.7312146997 + l0_4*2.844781386 + l0_5*2.4526730903 + l0_6*-1.9175165077 + l0_7*-0.7443755288 + l0_8*-3.1591419438 + l0_9*0.8441602697 + l0_10*1.1979484448 + l0_11*2.138098544 + l0_12*0.9274159536 + l0_13*-2.1573448803 + l0_14*-3.7698356464)
l1_21 = PineActivationFunctionTanh(l0_0*5.187120117 + l0_1*-7.7525670576 + l0_2*1.9008346975 + l0_3*-1.2031603996 + l0_4*5.917669142 + l0_5*-3.1878682719 + l0_6*1.0311747828 + l0_7*-2.7529484612 + l0_8*-1.1165884578 + l0_9*2.5524942323 + l0_10*-0.38623241 + l0_11*3.7961317445 + l0_12*-6.128820883 + l0_13*-2.1470707709 + l0_14*2.0173792965)
l1_22 = PineActivationFunctionTanh(l0_0*-6.0241676562 + l0_1*0.7474455584 + l0_2*1.7435724844 + l0_3*0.8619835076 + l0_4*-0.1138406797 + l0_5*6.5979359352 + l0_6*1.6554154348 + l0_7*-3.7969458806 + l0_8*1.1139097376 + l0_9*-1.9588417 + l0_10*3.5123392221 + l0_11*9.4443103128 + l0_12*-7.4779291395 + l0_13*3.6975940671 + l0_14*8.5134262747)
l1_23 = PineActivationFunctionTanh(l0_0*-7.5486576471 + l0_1*-0.0281420865 + l0_2*-3.8586839454 + l0_3*-0.5648792233 + l0_4*-7.3927282026 + l0_5*-0.3857538046 + l0_6*-2.9779885698 + l0_7*4.0482279965 + l0_8*-1.1522499578 + l0_9*-4.1562500212 + l0_10*0.7813134307 + l0_11*-1.7582667612 + l0_12*1.7071109988 + l0_13*6.9270873208 + l0_14*-4.5871357362)
l1_24 = PineActivationFunctionTanh(l0_0*-5.3603442228 + l0_1*-9.5350611629 + l0_2*1.6749984422 + l0_3*-0.6511065892 + l0_4*-0.8424823239 + l0_5*1.9946675213 + l0_6*-1.1264361638 + l0_7*0.3228676616 + l0_8*5.3562230396 + l0_9*-1.6678168952+ l0_10*1.2612580068 + l0_11*-3.5362671399 + l0_12*-9.3895191366 + l0_13*2.0169228673 + l0_14*-3.3813191557)
l1_25 = PineActivationFunctionTanh(l0_0*1.1362866429 + l0_1*-1.8960071702 + l0_2*5.7047307243 + l0_3*-1.6049785053 + l0_4*-4.8353898931 + l0_5*-1.4865381145 + l0_6*-0.2846893475 + l0_7*2.2322095997 + l0_8*2.0930488668 + l0_9*1.7141411002 + l0_10*-3.4106032176 + l0_11*3.0593289612 + l0_12*-5.0894813904 + l0_13*-0.5316299133 + l0_14*0.4705265416)
l1_26 = PineActivationFunctionTanh(l0_0*-0.9401400975 + l0_1*-0.9136086957 + l0_2*-3.3808688582 + l0_3*4.7200776773 + l0_4*3.686296919 + l0_5*14.2133723935 + l0_6*1.5652940954 + l0_7*-0.2921139433 + l0_8*1.0244504511 + l0_9*-7.6918299134 + l0_10*-0.594936135 + l0_11*-1.4559914156 + l0_12*2.8056435224 + l0_13*2.6103905733 + l0_14*2.3412348872)
l1_27 = PineActivationFunctionTanh(l0_0*1.1573980186 + l0_1*2.9593661909 + l0_2*0.4512594325 + l0_3*-0.9357210858 + l0_4*-1.2445804495 + l0_5*4.2716471631 + l0_6*1.5167912375 + l0_7*1.5026853293 + l0_8*1.3574772038 + l0_9*-1.9754386842 + l0_10*6.727671436 + l0_11*8.0145772889 + l0_12*7.3108970663 + l0_13*-2.5005627841 + l0_14*8.9604502277)
l1_28 = PineActivationFunctionTanh(l0_0*6.3576350212 + l0_1*-2.9731672725 + l0_2*-2.7763558082 + l0_3*-3.7902984555 + l0_4*-1.0065574585 + l0_5*-0.7011836061 + l0_6*-1.0298068578 + l0_7*1.201007784 + l0_8*-0.7835862254 + l0_9*-3.9863597435 + l0_10*6.7851825502 + l0_11*1.1120256721 + l0_12*-2.263287351 + l0_13*1.8314374104 + l0_14*-2.279102097)
l1_29 = PineActivationFunctionTanh(l0_0*-7.8741911036 + l0_1*-5.3370618518 + l0_2*11.9153868964 + l0_3*-4.1237170553 + l0_4*2.9491152758 + l0_5*1.0317132502 + l0_6*2.2992199883 + l0_7*-2.0250502364 + l0_8*-11.0785995839 + l0_9*-6.3615588554 + l0_10*-1.1687644976 + l0_11*6.3323478015 + l0_12*6.0195076962 + l0_13*-2.8972208702 + l0_14*3.6107747183)
 
l2_0 = PineActivationFunctionTanh(l1_0*-0.590546797 + l1_1*0.6608304658 + l1_2*-0.3358268839 + l1_3*-0.748530283 + l1_4*-0.333460383 + l1_5*-0.3409307681 + l1_6*0.1916558198 + l1_7*-0.1200399453 + l1_8*-0.5166151854 + l1_9*-0.8537164676 +l1_10*-0.0214448647 + l1_11*-0.553290271 + l1_12*-1.2333302892 + l1_13*-0.8321813811 + l1_14*-0.4527761741 + l1_15*0.9012545631 + l1_16*0.415853215 + l1_17*0.1270548319 + l1_18*0.2000460279 + l1_19*-0.1741942671 + l1_20*0.419830522 + l1_21*-0.059839291 + l1_22*-0.3383001769 + l1_23*0.1617814073 + l1_24*0.3071848006 + l1_25*-0.3191182045 + l1_26*-0.4981831822 + l1_27*-1.467478375 + l1_28*-0.1676432563 + l1_29*1.2574849126)
l2_1 = PineActivationFunctionTanh(l1_0*-0.5514235841 + l1_1*0.4759190049 + l1_2*0.2103576983 + l1_3*-0.4754377924 + l1_4*-0.2362941295 + l1_5*0.1155082119 + l1_6*0.7424215794 + l1_7*-0.3674198672 + l1_8*0.8401574461 + l1_9*0.6096563193 + l1_10*0.7437935674 + l1_11*-0.4898638101 + l1_12*-0.4168668092 + l1_13*-0.0365111095 + l1_14*-0.342675224 + l1_15*0.1870268765 + l1_16*-0.5843050987 + l1_17*-0.4596547471 + l1_18*0.452188522 + l1_19*-0.6737126684 + l1_20*0.6876072741 + l1_21*-0.8067776704 + l1_22*0.7592979467 + l1_23*-0.0768239468 + l1_24*0.370536097 + l1_25*-0.4363884671 + l1_26*-0.419285676 + l1_27*0.4380251141 + l1_28*0.0822528948 + l1_29*-0.2333910809)
l2_2 = PineActivationFunctionTanh(l1_0*-0.3306539521 + l1_1*-0.9382247194 + l1_2*0.0746711276 + l1_3*-0.3383838985 + l1_4*-0.0683232217 + l1_5*-0.2112358049 + l1_6*-0.9079234054 + l1_7*0.4898595603 + l1_8*-0.2039825863 + l1_9*1.0870698641+ l1_10*-1.1752901237 + l1_11*1.1406403923 + l1_12*-0.6779626786 + l1_13*0.4281048906 + l1_14*-0.6327670055 + l1_15*-0.1477678844 + l1_16*0.2693637584 + l1_17*0.7250738509 + l1_18*0.7905904504 + l1_19*-1.6417250883 + l1_20*-0.2108095534 +l1_21*-0.2698557472 + l1_22*-0.2433656685 + l1_23*-0.6289943273 + l1_24*0.436428207 + l1_25*-0.8243825184 + l1_26*-0.8583496686 + l1_27*0.0983131026 + l1_28*-0.4107462518 + l1_29*0.5641683087)
l2_3 = PineActivationFunctionTanh(l1_0*1.7036869992 + l1_1*-0.6683507666 + l1_2*0.2589197112 + l1_3*0.032841148 + l1_4*-0.4454796342 + l1_5*-0.6196149423 + l1_6*-0.1073622976 + l1_7*-0.1926393101 + l1_8*1.5280232458 + l1_9*-0.6136527036 +l1_10*-1.2722934357 + l1_11*0.2888655811 + l1_12*-1.4338638512 + l1_13*-1.1903556863 + l1_14*-1.7659663905 + l1_15*0.3703086867 + l1_16*1.0409140889 + l1_17*0.0167382209 + l1_18*0.6045646461 + l1_19*4.2388788116 + l1_20*1.4399738234 + l1_21*0.3308571935 + l1_22*1.4501137667 + l1_23*0.0426123724 + l1_24*-0.708479795 + l1_25*-1.2100800732 + l1_26*-0.5536278651 + l1_27*1.3547250573 + l1_28*1.2906250286 + l1_29*0.0596007114)
l2_4 = PineActivationFunctionTanh(l1_0*-0.462165126 + l1_1*-1.0996742176 + l1_2*1.0928262999 + l1_3*1.806407067 + l1_4*0.9289147669 + l1_5*0.8069022793 + l1_6*0.2374237802 + l1_7*-2.7143979019 + l1_8*-2.7779203877 + l1_9*0.214383903 + l1_10*-1.3111536623 + l1_11*-2.3148813568 + l1_12*-2.4755355804 + l1_13*-0.6819733236 + l1_14*0.4425615226 + l1_15*-0.1298218043 + l1_16*-1.1744832824 + l1_17*-0.395194848 + l1_18*-0.2803397703 + l1_19*-0.4505071197 + l1_20*-0.8934956598 + l1_21*3.3232916348 + l1_22*-1.7359534851 + l1_23*3.8540421743 + l1_24*1.4424032523 + l1_25*0.2639823693 + l1_26*0.3597053634 + l1_27*-1.0470693728 + l1_28*1.4133480357 + l1_29*0.6248098695)
l2_5 = PineActivationFunctionTanh(l1_0*0.2215807411 + l1_1*-0.5628295071 + l1_2*-0.8795982905 + l1_3*0.9101585104 + l1_4*-1.0176831976 + l1_5*-0.0728884401 + l1_6*0.6676331658 + l1_7*-0.7342174108 + l1_8*9.4428E-4 + l1_9*0.6439774272 + l1_10*-0.0345236026 + l1_11*0.5830977027 + l1_12*-0.4058921837 + l1_13*-0.3991888077 + l1_14*-1.0090426973 + l1_15*-0.9324780698 + l1_16*-0.0888749165 + l1_17*0.2466351736 + l1_18*0.4993304601 + l1_19*-1.115408696 + l1_20*0.9914246705 + l1_21*0.9687743445 + l1_22*0.1117130875 + l1_23*0.7825109733 + l1_24*0.2217023612 + l1_25*0.3081256411 + l1_26*-0.1778007966 + l1_27*-0.3333287743 + l1_28*1.0156352461 + l1_29*-0.1456257813)
l2_6 = PineActivationFunctionTanh(l1_0*-0.5461783383 + l1_1*0.3246015999 + l1_2*0.1450605434 + l1_3*-1.3179944349 + l1_4*-1.5481775261 + l1_5*-0.679685633 + l1_6*-0.9462335139 + l1_7*-0.6462399371 + l1_8*0.0991658683 + l1_9*0.1612892194 +l1_10*-1.037660602 + l1_11*-0.1044778824 + l1_12*0.8309203243 + l1_13*0.7714766458 + l1_14*0.2566767663 + l1_15*0.8649416329 + l1_16*-0.5847461285 + l1_17*-0.6393969272 + l1_18*0.8014049359 + l1_19*0.2279568228 + l1_20*1.0565217821 + l1_21*0.134738029 + l1_22*0.3420395576 + l1_23*-0.2417397219 + l1_24*0.3083072038 + l1_25*0.6761739059 + l1_26*-0.4653817053 + l1_27*-1.0634057566 + l1_28*-0.5658892281 + l1_29*-0.6947283681)
l2_7 = PineActivationFunctionTanh(l1_0*-0.5450410944 + l1_1*0.3912849372 + l1_2*-0.4118641117 + l1_3*0.7124695074 + l1_4*-0.7510266122 + l1_5*1.4065673913 + l1_6*0.9870731545 + l1_7*-0.2609363107 + l1_8*-0.3583639958 + l1_9*0.5436375706 +l1_10*0.4572450099 + l1_11*-0.4651538878 + l1_12*-0.2180218212 + l1_13*0.5241262959 + l1_14*-0.8529323253 + l1_15*-0.4200378937 + l1_16*0.4997885721 + l1_17*-1.1121528189 + l1_18*0.5992411048 + l1_19*-1.0263270781 + l1_20*-1.725160642 + l1_21*-0.2653995722 + l1_22*0.6996703032 + l1_23*0.348549086 + l1_24*0.6522482482 + l1_25*-0.7931928436 + l1_26*-0.5107994359 + l1_27*0.0509642698 + l1_28*0.8711187423 + l1_29*0.8999449627)
l2_8 = PineActivationFunctionTanh(l1_0*-0.7111081522 + l1_1*0.4296245062 + l1_2*-2.0720732038 + l1_3*-0.4071818684 + l1_4*1.0632721681 + l1_5*0.8463224325 + l1_6*-0.6083948423 + l1_7*1.1827669608 + l1_8*-0.9572307844 + l1_9*-0.9080517673 + l1_10*-0.0479029057 + l1_11*-1.1452853213 + l1_12*0.2884352688 + l1_13*0.1767851586 + l1_14*-1.089314461 + l1_15*1.2991763966 + l1_16*1.6236630806 + l1_17*-0.7720263697 + l1_18*-0.5011541755 + l1_19*-2.3919413568 + l1_20*0.0084018338 + l1_21*0.9975216139 + l1_22*0.4193541029 + l1_23*1.4623834571 + l1_24*-0.6253069691 + l1_25*0.6119677341 + l1_26*0.5423948388 + l1_27*1.0022450377 + l1_28*-1.2392984069 + l1_29*1.5021529822)
 
l3_0 = PineActivationFunctionTanh(l2_0*0.3385061186 + l2_1*0.6218531956 + l2_2*-0.7790340983 + l2_3*0.1413078332 + l2_4*0.1857010624 + l2_5*-0.1769456351 + l2_6*-0.3242337911 + l2_7*-0.503944883 + l2_8*0.1540568869)
 
hline(0, title="base line")
bgcolor(l3_0 > 0 ? green : red, transp=20)
plot(l3_0, color=aqua, title="prediction")

@sirolf2009 Thank you. Interesting indicator.
1) It repaints the current candle. Do you take action on candle close?
2) How do you suggest to take away the choppiness during sideways periods?
Thank you in advance!
+8 Phản hồi
Really interesting!
Thank you very much for this!
Great job!
Phản hồi
Hi, what is the loss-function you used? I'm currenctly testing some stacked LSTM layers.
+7 Phản hồi
does the indis repaint ??
+4 Phản hồi
Nicee....
+2 Phản hồi
Thank you very much. You're indeed a genius!
+1 Phản hồi
how to apply this indicator to chart
+7 Phản hồi
amazing thanks
Phản hồi
You're a genius! Thank you for all your support.
+1 Phản hồi
Trang chủ Bộ lọc cổ phiếu Bộ lọc Forex Bộ lọc Tiền điện tử Lịch kinh tế Cách thức hoạt động Tính năng Biểu đồ Giá Quy tắc Áp dụng Người điều hành Giải pháp cho Website & Nhà môi giới Widget Giải pháp biểu đồ Trung tâm Trợ giúp Yêu cầu tính năng Blog & Tin tức Câu hỏi thường gặp Wiki Twitter
Hồ sơ Tùy chỉnh Hồ sơ Tài khoản và Thanh toán Đồng TradingView Ticket Hỗ trợ của tôi Trung tâm Trợ giúp Các ý tưởng đã đăng Người theo dõi Đang theo dõi Tin nhắn riêng Trò chuyện Đăng xuất