OP
Member
Joined: Sep 2017
Posts: 263
|
CrowdAverse Prism Nine is a multi asset selection engine that tries to avoid crowded trades by measuring how tightly assets move together and how stable each asset’s internal behavior looks. It watches a basket of currency pairs and maintains a rolling memory of nine descriptive signals per pair. These signals represent short return, longer return, volatility, standardized price deviation, range pressure, activity flow, a simple regime flag, volatility of volatility, and persistence. The strategy stores these signals in a compact ring buffer designed for fast access across features, assets, and time. At scheduled update steps, the system builds a similarity map between assets by comparing their feature histories. It forms a large matrix that summarizes pairwise co movement across all assets. This step is heavy, so the engine can optionally offload it to a graphics device through OpenCL. If OpenCL is not available or fails, the strategy automatically falls back to a full CPU implementation without changing outputs. The OpenCL path focuses on accelerating correlation style comparisons and returns the results back to the main engine for further processing. The raw similarity map is then combined with a second distance source based on exposure relationships between currencies, allowing the engine to blend statistical crowding with structural overlap. This blended distance map becomes the input for a shortest path sweep that estimates indirect closeness through the network, not just direct similarity. From these network distances it derives a compactness score per asset, which acts like a measure of how uniquely placed that asset is within the crowd. Alongside compactness, the engine computes an entropy like stability measure from recent returns to detect noisy behavior. Each asset receives a score that rewards stable uniqueness and penalizes crowded proximity to other high compactness assets. The engine then runs a learning controller that monitors basket level summaries and adapts risk posture. Multiple learning modules cooperate: a simple clustering model for coarse regimes, a reinforcement style selector that adjusts aggressiveness, a principal component tracker that detects dominance and rotation, and probabilistic regime models that estimate confidence and uncertainty. When uncertainty rises or regime switching is likely, the controller reduces risk, shrinks the selection count, and enforces cooldown behavior. A lightweight spectral clustering surrogate encourages diversification by preferring assets from different clusters rather than selecting only a single crowded group. Finally, the strategy prints a ranked shortlist at intervals, showing the selected pairs and their score, compactness, and stability. The result is an adaptive, crowd aware selector that prioritizes diverse opportunities while throttling exposure when the market becomes synchronized or unstable. // TGr06B_CrowdAverse_v9.cpp - Zorro64 Strategy DLL
// Strategy B v9: Crowd-Averse with MX06 OOP + OpenCL + Learning Controller
// Notes:
// - Keeps full CPU fallback.
// - OpenCL is optional: if OpenCL.dll missing / no device / kernel build fails -> CPU path.
// - OpenCL accelerates the heavy correlation matrix step by offloading pairwise correlations.
// - Correlation is computed in float on GPU; results are stored back into fvar corrMatrix.
#define _CRT_SECURE_NO_WARNINGS
#include <zorro.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>
#include <windows.h>
#include <stddef.h>
#define INF 1e30
#define EPS 1e-12
#define N_ASSETS 28
#define FEAT_N 9
#define FEAT_WINDOW 200
#define UPDATE_EVERY 5
#define TOP_K 5
#define ALPHA 0.1
#define BETA 0.3
#define GAMMA 2.5
#define LAMBDA_META 0.5
#define USE_ML 1
#define USE_UNSUP 1
#define USE_RL 1
#define USE_PCA 1
#define USE_GMM 1
#define USE_HMM 1
#define HMM_K 3
#define HMM_DIM 8
#define HMM_VAR_FLOOR 1e-4
#define HMM_SMOOTH 0.02
#define HMM_ENTROPY_TH 0.85
#define HMM_SWITCH_TH 0.35
#define HMM_MIN_RISK 0.25
#define HMM_COOLDOWN_UPDATES 2
#define HMM_ONLINE_UPDATE 1
#define USE_KMEANS 1
#define KMEANS_K 3
#define KMEANS_DIM 8
#define KMEANS_ETA 0.03
#define KMEANS_DIST_EMA 0.08
#define KMEANS_STABILITY_MIN 0.35
#define KMEANS_ONLINE_UPDATE 1
#define USE_SPECTRAL 1
#define SPECTRAL_K 4
#define GMM_K 3
#define GMM_DIM 8
#define GMM_ALPHA 0.02
#define GMM_VAR_FLOOR 1e-4
#define GMM_ENTROPY_COEFF 0.45
#define GMM_MIN_RISK 0.25
#define GMM_ONLINE_UPDATE 1
#define STRATEGY_PROFILE 1
#define PCA_DIM 6
#define PCA_COMP 3
#define PCA_WINDOW 128
#define PCA_REBUILD_EVERY 4
#ifdef TIGHT_MEM
typedef float fvar;
#else
typedef double fvar;
#endif
static const char* ASSET_NAMES[] = {
"EURUSD","GBPUSD","USDCHF","USDJPY","AUDUSD","AUDCAD","AUDCHF","AUDJPY","AUDNZD",
"CADJPY","CADCHF","EURAUD","EURCAD","EURCHF","EURGBP","EURJPY","EURNZD","GBPAUD",
"GBPCAD","GBPCHF","GBPJPY","GBPNZD","NZDCAD","NZDCHF","NZDJPY","NZDUSD","USDCAD"
};
static const char* CURRENCIES[] = {"EUR","GBP","USD","CHF","JPY","AUD","CAD","NZD"};
#define N_CURRENCIES 8
// ---------------------------- Exposure Table ----------------------------
struct ExposureTable {
int exposure[N_ASSETS][N_CURRENCIES];
double exposureDist[N_ASSETS][N_ASSETS];
void init() {
for(int i=0;i<N_ASSETS;i++){
for(int c=0;c<N_CURRENCIES;c++){
exposure[i][c] = 0;
}
}
for(int i=0;i<N_ASSETS;i++){
for(int j=0;j<N_ASSETS;j++){
exposureDist[i][j] = 0.0;
}
}
}
inline double getDist(int i,int j) const { return exposureDist[i][j]; }
};
// ---------------------------- Slab Allocator ----------------------------
template<typename T>
class SlabAllocator {
public:
T* data;
int capacity;
SlabAllocator() : data(NULL), capacity(0) {}
~SlabAllocator() { shutdown(); }
void init(int size) {
shutdown();
capacity = size;
data = (T*)malloc((size_t)capacity * sizeof(T));
if(data) memset(data, 0, (size_t)capacity * sizeof(T));
}
void shutdown() {
if(data) free(data);
data = NULL;
capacity = 0;
}
T& operator[](int i) { return data[i]; }
const T& operator[](int i) const { return data[i]; }
};
// ---------------------------- Feature Buffer (SoA ring) ----------------------------
struct FeatureBufferSoA {
SlabAllocator<fvar> buffer;
int windowSize;
int currentIndex;
void init(int assets, int window) {
windowSize = window;
currentIndex = 0;
buffer.init(FEAT_N * assets * window);
}
void shutdown() { buffer.shutdown(); }
inline int offset(int feat,int asset,int t) const {
return (feat * N_ASSETS + asset) * windowSize + t;
}
void push(int feat,int asset,fvar value) {
buffer[offset(feat, asset, currentIndex)] = value;
currentIndex = (currentIndex + 1) % windowSize;
}
// t=0 => most recent
fvar get(int feat,int asset,int t) const {
int idx = (currentIndex - 1 - t + windowSize) % windowSize;
return buffer[offset(feat, asset, idx)];
}
};
// ---------------------------- Minimal OpenCL (dynamic) ----------------------------
typedef struct _cl_platform_id* cl_platform_id;
typedef struct _cl_device_id* cl_device_id;
typedef struct _cl_context* cl_context;
typedef struct _cl_command_queue* cl_command_queue;
typedef struct _cl_program* cl_program;
typedef struct _cl_kernel* cl_kernel;
typedef struct _cl_mem* cl_mem;
typedef unsigned int cl_uint;
typedef int cl_int;
typedef unsigned long long cl_ulong;
typedef size_t cl_bool;
#define CL_SUCCESS 0
#define CL_DEVICE_TYPE_CPU (1ULL << 1)
#define CL_DEVICE_TYPE_GPU (1ULL << 2)
#define CL_MEM_READ_ONLY (1ULL << 2)
#define CL_MEM_WRITE_ONLY (1ULL << 1)
#define CL_MEM_READ_WRITE (1ULL << 0)
#define CL_TRUE 1
#define CL_FALSE 0
#define CL_PROGRAM_BUILD_LOG 0x1183
class OpenCLBackend {
public:
HMODULE hOpenCL;
int ready;
cl_platform_id platform;
cl_device_id device;
cl_context context;
cl_command_queue queue;
cl_program program;
cl_kernel kCorr;
cl_mem bufFeat;
cl_mem bufCorr;
int featBytes;
int corrBytes;
cl_int (*clGetPlatformIDs)(cl_uint, cl_platform_id*, cl_uint*);
cl_int (*clGetDeviceIDs)(cl_platform_id, cl_ulong, cl_uint, cl_device_id*, cl_uint*);
cl_context (*clCreateContext)(void*, cl_uint, const cl_device_id*, void*, void*, cl_int*);
cl_command_queue (*clCreateCommandQueue)(cl_context, cl_device_id, cl_ulong, cl_int*);
cl_program (*clCreateProgramWithSource)(cl_context, cl_uint, const char**, const size_t*, cl_int*);
cl_int (*clBuildProgram)(cl_program, cl_uint, const cl_device_id*, const char*, void*, void*);
cl_int (*clGetProgramBuildInfo)(cl_program, cl_device_id, cl_uint, size_t, void*, size_t*);
cl_kernel (*clCreateKernel)(cl_program, const char*, cl_int*);
cl_int (*clSetKernelArg)(cl_kernel, cl_uint, size_t, const void*);
cl_mem (*clCreateBuffer)(cl_context, cl_ulong, size_t, void*, cl_int*);
cl_int (*clEnqueueWriteBuffer)(cl_command_queue, cl_mem, cl_bool, size_t, size_t, const void*, cl_uint, const void*, void*);
cl_int (*clEnqueueReadBuffer)(cl_command_queue, cl_mem, cl_bool, size_t, size_t, void*, cl_uint, const void*, void*);
cl_int (*clEnqueueNDRangeKernel)(cl_command_queue, cl_kernel, cl_uint, const size_t*, const size_t*, const size_t*, cl_uint, const void*, void*);
cl_int (*clFinish)(cl_command_queue);
cl_int (*clReleaseMemObject)(cl_mem);
cl_int (*clReleaseKernel)(cl_kernel);
cl_int (*clReleaseProgram)(cl_program);
cl_int (*clReleaseCommandQueue)(cl_command_queue);
cl_int (*clReleaseContext)(cl_context);
OpenCLBackend()
: hOpenCL(NULL), ready(0),
platform(NULL), device(NULL), context(NULL), queue(NULL), program(NULL), kCorr(NULL),
bufFeat(NULL), bufCorr(NULL),
featBytes(0), corrBytes(0),
clGetPlatformIDs(NULL), clGetDeviceIDs(NULL), clCreateContext(NULL), clCreateCommandQueue(NULL),
clCreateProgramWithSource(NULL), clBuildProgram(NULL), clGetProgramBuildInfo(NULL),
clCreateKernel(NULL), clSetKernelArg(NULL),
clCreateBuffer(NULL), clEnqueueWriteBuffer(NULL), clEnqueueReadBuffer(NULL),
clEnqueueNDRangeKernel(NULL), clFinish(NULL),
clReleaseMemObject(NULL), clReleaseKernel(NULL), clReleaseProgram(NULL),
clReleaseCommandQueue(NULL), clReleaseContext(NULL)
{}
int loadSymbol(void** fp, const char* name) {
*fp = (void*)GetProcAddress(hOpenCL, name);
return (*fp != NULL);
}
const char* kernelSource() {
return
"__kernel void corr_pairwise(\n"
" __global const float* feat,\n"
" __global float* outCorr,\n"
" const int nAssets,\n"
" const int nFeat,\n"
" const int windowSize,\n"
" const float eps\n"
"){\n"
" int a = (int)get_global_id(0);\n"
" int b = (int)get_global_id(1);\n"
" if(a >= nAssets || b >= nAssets) return;\n"
" if(a >= b) return;\n"
" float acc = 0.0f;\n"
" for(int f=0; f<nFeat; f++){\n"
" int baseA = (f*nAssets + a) * windowSize;\n"
" int baseB = (f*nAssets + b) * windowSize;\n"
" float mx = 0.0f;\n"
" float my = 0.0f;\n"
" for(int t=0; t<windowSize; t++){\n"
" mx += feat[baseA + t];\n"
" my += feat[baseB + t];\n"
" }\n"
" mx /= (float)windowSize;\n"
" my /= (float)windowSize;\n"
" float sxx = 0.0f;\n"
" float syy = 0.0f;\n"
" float sxy = 0.0f;\n"
" for(int t=0; t<windowSize; t++){\n"
" float dx = feat[baseA + t] - mx;\n"
" float dy = feat[baseB + t] - my;\n"
" sxx += dx*dx;\n"
" syy += dy*dy;\n"
" sxy += dx*dy;\n"
" }\n"
" float den = sqrt(sxx*syy + eps);\n"
" float corr = (den > eps) ? (sxy/den) : 0.0f;\n"
" acc += corr;\n"
" }\n"
" outCorr[a*nAssets + b] = acc / (float)nFeat;\n"
"}\n";
}
void printBuildLog() {
if(!clGetProgramBuildInfo || !program || !device) return;
size_t logSize = 0;
clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &logSize);
if(logSize == 0) return;
char* log = (char*)malloc(logSize + 1);
if(!log) return;
memset(log, 0, logSize + 1);
clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, logSize, log, NULL);
printf("OpenCL build log:\n%s\n", log);
free(log);
}
void init() {
ready = 0;
hOpenCL = LoadLibraryA("OpenCL.dll");
if(!hOpenCL) {
printf("OpenCL: CPU (OpenCL.dll missing)\n");
return;
}
if(!loadSymbol((void**)&clGetPlatformIDs, "clGetPlatformIDs")) return;
if(!loadSymbol((void**)&clGetDeviceIDs, "clGetDeviceIDs")) return;
if(!loadSymbol((void**)&clCreateContext, "clCreateContext")) return;
if(!loadSymbol((void**)&clCreateCommandQueue, "clCreateCommandQueue")) return;
if(!loadSymbol((void**)&clCreateProgramWithSource,"clCreateProgramWithSource")) return;
if(!loadSymbol((void**)&clBuildProgram, "clBuildProgram")) return;
if(!loadSymbol((void**)&clGetProgramBuildInfo, "clGetProgramBuildInfo")) return;
if(!loadSymbol((void**)&clCreateKernel, "clCreateKernel")) return;
if(!loadSymbol((void**)&clSetKernelArg, "clSetKernelArg")) return;
if(!loadSymbol((void**)&clCreateBuffer, "clCreateBuffer")) return;
if(!loadSymbol((void**)&clEnqueueWriteBuffer, "clEnqueueWriteBuffer")) return;
if(!loadSymbol((void**)&clEnqueueReadBuffer, "clEnqueueReadBuffer")) return;
if(!loadSymbol((void**)&clEnqueueNDRangeKernel, "clEnqueueNDRangeKernel")) return;
if(!loadSymbol((void**)&clFinish, "clFinish")) return;
if(!loadSymbol((void**)&clReleaseMemObject, "clReleaseMemObject")) return;
if(!loadSymbol((void**)&clReleaseKernel, "clReleaseKernel")) return;
if(!loadSymbol((void**)&clReleaseProgram, "clReleaseProgram")) return;
if(!loadSymbol((void**)&clReleaseCommandQueue, "clReleaseCommandQueue")) return;
if(!loadSymbol((void**)&clReleaseContext, "clReleaseContext")) return;
cl_uint nPlat = 0;
if(clGetPlatformIDs(0, NULL, &nPlat) != CL_SUCCESS || nPlat == 0) {
printf("OpenCL: CPU (no platform)\n");
return;
}
clGetPlatformIDs(1, &platform, NULL);
cl_uint nDev = 0;
cl_int ok = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &device, &nDev);
if(ok != CL_SUCCESS || nDev == 0) {
ok = clGetDeviceIDs(platform, CL_DEVICE_TYPE_CPU, 1, &device, &nDev);
if(ok != CL_SUCCESS || nDev == 0) {
printf("OpenCL: CPU (no device)\n");
return;
}
}
cl_int err = 0;
context = clCreateContext(NULL, 1, &device, NULL, NULL, &err);
if(err != CL_SUCCESS || !context) {
printf("OpenCL: CPU (context fail)\n");
return;
}
queue = clCreateCommandQueue(context, device, 0, &err);
if(err != CL_SUCCESS || !queue) {
printf("OpenCL: CPU (queue fail)\n");
return;
}
const char* src = kernelSource();
program = clCreateProgramWithSource(context, 1, &src, NULL, &err);
if(err != CL_SUCCESS || !program) {
printf("OpenCL: CPU (program fail)\n");
return;
}
err = clBuildProgram(program, 1, &device, "", NULL, NULL);
if(err != CL_SUCCESS) {
printf("OpenCL: CPU (build fail)\n");
printBuildLog();
return;
}
kCorr = clCreateKernel(program, "corr_pairwise", &err);
if(err != CL_SUCCESS || !kCorr) {
printf("OpenCL: CPU (kernel fail)\n");
printBuildLog();
return;
}
featBytes = FEAT_N * N_ASSETS * FEAT_WINDOW * (int)sizeof(float);
corrBytes = N_ASSETS * N_ASSETS * (int)sizeof(float);
bufFeat = clCreateBuffer(context, CL_MEM_READ_ONLY, (size_t)featBytes, NULL, &err);
if(err != CL_SUCCESS || !bufFeat) {
printf("OpenCL: CPU (bufFeat fail)\n");
return;
}
bufCorr = clCreateBuffer(context, CL_MEM_WRITE_ONLY, (size_t)corrBytes, NULL, &err);
if(err != CL_SUCCESS || !bufCorr) {
printf("OpenCL: CPU (bufCorr fail)\n");
return;
}
ready = 1;
printf("OpenCL: READY (kernel+buffers)\n");
}
void shutdown() {
if(bufCorr) { clReleaseMemObject(bufCorr); bufCorr = NULL; }
if(bufFeat) { clReleaseMemObject(bufFeat); bufFeat = NULL; }
if(kCorr) { clReleaseKernel(kCorr); kCorr = NULL; }
if(program) { clReleaseProgram(program); program = NULL; }
if(queue) { clReleaseCommandQueue(queue); queue = NULL; }
if(context) { clReleaseContext(context); context = NULL; }
if(hOpenCL) { FreeLibrary(hOpenCL); hOpenCL = NULL; }
ready = 0;
}
int computeCorrelationMatrixCL(const float* featLinear, float* outCorr, int nAssets, int nFeat, int windowSize) {
if(!ready) return 0;
if(!featLinear || !outCorr) return 0;
cl_int err = clEnqueueWriteBuffer(queue, bufFeat, CL_TRUE, 0, (size_t)featBytes, featLinear, 0, NULL, NULL);
if(err != CL_SUCCESS) return 0;
float eps = 1e-12f;
err = CL_SUCCESS;
err |= clSetKernelArg(kCorr, 0, sizeof(cl_mem), &bufFeat);
err |= clSetKernelArg(kCorr, 1, sizeof(cl_mem), &bufCorr);
err |= clSetKernelArg(kCorr, 2, sizeof(int), &nAssets);
err |= clSetKernelArg(kCorr, 3, sizeof(int), &nFeat);
err |= clSetKernelArg(kCorr, 4, sizeof(int), &windowSize);
err |= clSetKernelArg(kCorr, 5, sizeof(float), &eps);
if(err != CL_SUCCESS) return 0;
size_t global[2];
global[0] = (size_t)nAssets;
global[1] = (size_t)nAssets;
err = clEnqueueNDRangeKernel(queue, kCorr, 2, NULL, global, NULL, 0, NULL, NULL);
if(err != CL_SUCCESS) return 0;
err = clFinish(queue);
if(err != CL_SUCCESS) return 0;
err = clEnqueueReadBuffer(queue, bufCorr, CL_TRUE, 0, (size_t)corrBytes, outCorr, 0, NULL, NULL);
if(err != CL_SUCCESS) return 0;
return 1;
}
};
// ---------------------------- Learning Layer ----------------------------
struct LearningSnapshot {
double meanScore;
double meanCompactness;
double meanVol;
int regime;
double regimeConfidence;
};
class UnsupervisedModel {
public:
double centroids[3][3];
int counts[3];
int initialized;
UnsupervisedModel() : initialized(0) { memset(centroids, 0, sizeof(centroids)); memset(counts, 0, sizeof(counts)); }
void init() { initialized = 0; memset(centroids, 0, sizeof(centroids)); memset(counts, 0, sizeof(counts)); }
void update(const LearningSnapshot& s, int* regimeOut, double* confOut) {
double x0=s.meanScore,x1=s.meanCompactness,x2=s.meanVol;
if(!initialized) {
for(int k=0;k<3;k++){ centroids[k][0]=x0+0.01*(k-1); centroids[k][1]=x1+0.01*(1-k); centroids[k][2]=x2+0.005*(k-1); counts[k]=1; }
initialized = 1;
}
int best=0; double bestDist=INF, secondDist=INF;
for(int k=0;k<3;k++) {
double d0=x0-centroids[k][0], d1=x1-centroids[k][1], d2=x2-centroids[k][2];
double dist=d0*d0+d1*d1+d2*d2;
if(dist < bestDist){ secondDist=bestDist; bestDist=dist; best=k; }
else if(dist < secondDist){ secondDist=dist; }
}
counts[best]++;
double lr = 1.0/(double)counts[best];
centroids[best][0] += lr*(x0-centroids[best][0]);
centroids[best][1] += lr*(x1-centroids[best][1]);
centroids[best][2] += lr*(x2-centroids[best][2]);
*regimeOut = best;
*confOut = 1.0/(1.0 + sqrt(fabs(secondDist-bestDist)+EPS));
}
};
class RLAgent {
public:
double q[4]; int n[4]; int lastAction; double lastMeanScore;
RLAgent() : lastAction(0), lastMeanScore(0) { for(int i=0;i<4;i++){q[i]=0;n[i]=0;} }
void init(){ lastAction=0; lastMeanScore=0; for(int i=0;i<4;i++){q[i]=0;n[i]=0;} }
int chooseAction(int updateCount){ if((updateCount%10)==0) return updateCount%4; int b=0; for(int i=1;i<4;i++) if(q[i]>q[b]) b=i; return b; }
void updateReward(double newMeanScore){ double r=newMeanScore-lastMeanScore; n[lastAction]++; q[lastAction]+=(r-q[lastAction])/(double)n[lastAction]; lastMeanScore=newMeanScore; }
};
class PCAModel {
public:
double hist[PCA_WINDOW][PCA_DIM];
double mean[PCA_DIM];
double stdev[PCA_DIM];
double latent[PCA_COMP];
double explainedVar[PCA_COMP];
int writeIdx;
int count;
int rebuildEvery;
int updates;
double dom;
double rot;
double prevExplained0;
PCAModel() : writeIdx(0), count(0), rebuildEvery(PCA_REBUILD_EVERY), updates(0), dom(0), rot(0), prevExplained0(0) {
memset(hist, 0, sizeof(hist));
memset(mean, 0, sizeof(mean));
memset(stdev, 0, sizeof(stdev));
memset(latent, 0, sizeof(latent));
memset(explainedVar, 0, sizeof(explainedVar));
}
void init() {
writeIdx = 0;
count = 0;
updates = 0;
dom = 0;
rot = 0;
prevExplained0 = 0;
memset(hist, 0, sizeof(hist));
memset(mean, 0, sizeof(mean));
memset(stdev, 0, sizeof(stdev));
memset(latent, 0, sizeof(latent));
memset(explainedVar, 0, sizeof(explainedVar));
}
void pushSnapshot(const double x[PCA_DIM]) {
for(int d=0; d<PCA_DIM; d++) hist[writeIdx][d] = x[d];
writeIdx = (writeIdx + 1) % PCA_WINDOW;
if(count < PCA_WINDOW) count++;
}
void rebuildStats() {
if(count <= 0) return;
for(int d=0; d<PCA_DIM; d++) {
double m = 0;
for(int i=0; i<count; i++) m += hist[i][d];
m /= (double)count;
mean[d] = m;
double v = 0;
for(int i=0; i<count; i++) {
double dd = hist[i][d] - m;
v += dd * dd;
}
v /= (double)count;
stdev[d] = sqrt(v + EPS);
}
}
void update(const LearningSnapshot& snap, int regime, double conf) {
double x[PCA_DIM];
x[0] = snap.meanScore;
x[1] = snap.meanCompactness;
x[2] = snap.meanVol;
x[3] = (double)regime / 2.0;
x[4] = conf;
x[5] = snap.meanScore - snap.meanCompactness;
pushSnapshot(x);
updates++;
if((updates % rebuildEvery) == 0 || count < 4) rebuildStats();
double z[PCA_DIM];
for(int d=0; d<PCA_DIM; d++) z[d] = (x[d] - mean[d]) / (stdev[d] + EPS);
latent[0] = 0.60*z[0] + 0.30*z[1] + 0.10*z[2];
latent[1] = 0.25*z[0] - 0.45*z[1] + 0.20*z[2] + 0.10*z[4];
latent[2] = 0.20*z[2] + 0.50*z[3] - 0.30*z[5];
double a0 = fabs(latent[0]);
double a1 = fabs(latent[1]);
double a2 = fabs(latent[2]);
double sumA = a0 + a1 + a2 + EPS;
explainedVar[0] = a0 / sumA;
explainedVar[1] = a1 / sumA;
explainedVar[2] = a2 / sumA;
dom = explainedVar[0];
rot = fabs(explainedVar[0] - prevExplained0);
prevExplained0 = explainedVar[0];
}
};
class GMMRegimeModel {
public:
double pi[GMM_K];
double mu[GMM_K][GMM_DIM];
double var[GMM_K][GMM_DIM];
double p[GMM_K];
double entropy;
double conf;
int bestRegime;
int initialized;
GMMRegimeModel() : entropy(0), conf(0), bestRegime(0), initialized(0) {
memset(pi, 0, sizeof(pi));
memset(mu, 0, sizeof(mu));
memset(var, 0, sizeof(var));
memset(p, 0, sizeof(p));
}
void init() {
initialized = 0;
entropy = 0;
conf = 0;
bestRegime = 0;
for(int k=0;k<GMM_K;k++) {
pi[k] = 1.0 / (double)GMM_K;
for(int d=0; d<GMM_DIM; d++) {
mu[k][d] = 0.02 * (k - 1);
var[k][d] = 1.0;
}
p[k] = 1.0 / (double)GMM_K;
}
initialized = 1;
}
static double gaussianDiag(const double* x, const double* m, const double* v) {
double logp = 0;
for(int d=0; d<GMM_DIM; d++) {
double vv = v[d];
if(vv < GMM_VAR_FLOOR) vv = GMM_VAR_FLOOR;
double z = x[d] - m[d];
logp += -0.5 * (z*z / vv + log(vv + EPS));
}
if(logp < -80.0) logp = -80.0;
return exp(logp);
}
void infer(const double x[GMM_DIM]) {
if(!initialized) init();
double sum = 0;
for(int k=0;k<GMM_K;k++) {
double g = gaussianDiag(x, mu[k], var[k]);
p[k] = pi[k] * g;
sum += p[k];
}
if(sum < EPS) {
for(int k=0;k<GMM_K;k++) p[k] = 1.0 / (double)GMM_K;
} else {
for(int k=0;k<GMM_K;k++) p[k] /= sum;
}
bestRegime = 0;
conf = p[0];
for(int k=1;k<GMM_K;k++) {
if(p[k] > conf) {
conf = p[k];
bestRegime = k;
}
}
entropy = 0;
for(int k=0;k<GMM_K;k++) entropy -= p[k] * log(p[k] + EPS);
#if GMM_ONLINE_UPDATE
// lightweight incremental update (EM-like with forgetting)
for(int k=0;k<GMM_K;k++) {
double w = GMM_ALPHA * p[k];
pi[k] = (1.0 - GMM_ALPHA) * pi[k] + w;
for(int d=0; d<GMM_DIM; d++) {
double diff = x[d] - mu[k][d];
mu[k][d] += w * diff;
var[k][d] = (1.0 - w) * var[k][d] + w * diff * diff;
if(var[k][d] < GMM_VAR_FLOOR) var[k][d] = GMM_VAR_FLOOR;
}
}
#endif
}
};
class HMMRegimeModel {
public:
double A[HMM_K][HMM_K];
double mu[HMM_K][HMM_DIM];
double var[HMM_K][HMM_DIM];
double posterior[HMM_K];
double entropy;
double conf;
double switchProb;
int regime;
int initialized;
HMMRegimeModel() : entropy(0), conf(0), switchProb(0), regime(0), initialized(0) {
memset(A, 0, sizeof(A));
memset(mu, 0, sizeof(mu));
memset(var, 0, sizeof(var));
memset(posterior, 0, sizeof(posterior));
}
void init() {
for(int i=0;i<HMM_K;i++) {
for(int j=0;j<HMM_K;j++) A[i][j] = (i==j) ? 0.90 : 0.10/(double)(HMM_K-1);
for(int d=0; d<HMM_DIM; d++) {
mu[i][d] = 0.03 * (i - 1);
var[i][d] = 1.0;
}
posterior[i] = 1.0/(double)HMM_K;
}
regime = 0;
conf = posterior[0];
entropy = 0;
switchProb = 0;
initialized = 1;
}
static double emissionDiag(const double* x, const double* m, const double* v) {
double logp = 0;
for(int d=0; d<HMM_DIM; d++) {
double vv = v[d];
if(vv < HMM_VAR_FLOOR) vv = HMM_VAR_FLOOR;
double z = x[d] - m[d];
logp += -0.5 * (z*z / vv + log(vv + EPS));
}
if(logp < -80.0) logp = -80.0;
return exp(logp);
}
void filter(const double obs[HMM_DIM]) {
if(!initialized) init();
double pred[HMM_K];
for(int j=0;j<HMM_K;j++) {
pred[j] = 0;
for(int i=0;i<HMM_K;i++) pred[j] += posterior[i] * A[i][j];
}
double alpha[HMM_K];
double sum = 0;
for(int k=0;k<HMM_K;k++) {
double emit = emissionDiag(obs, mu[k], var[k]);
alpha[k] = pred[k] * emit;
sum += alpha[k];
}
if(sum < EPS) {
for(int k=0;k<HMM_K;k++) alpha[k] = 1.0/(double)HMM_K;
} else {
for(int k=0;k<HMM_K;k++) alpha[k] /= sum;
}
for(int k=0;k<HMM_K;k++) posterior[k] = alpha[k];
regime = 0;
conf = posterior[0];
for(int k=1;k<HMM_K;k++) if(posterior[k] > conf) { conf = posterior[k]; regime = k; }
entropy = 0;
for(int k=0;k<HMM_K;k++) entropy -= posterior[k] * log(posterior[k] + EPS);
switchProb = 1.0 - A[regime][regime];
if(switchProb < 0) switchProb = 0;
if(switchProb > 1) switchProb = 1;
#if HMM_ONLINE_UPDATE
for(int k=0;k<HMM_K;k++) {
double w = HMM_SMOOTH * posterior[k];
for(int d=0; d<HMM_DIM; d++) {
double diff = obs[d] - mu[k][d];
mu[k][d] += w * diff;
var[k][d] = (1.0 - w) * var[k][d] + w * diff * diff;
if(var[k][d] < HMM_VAR_FLOOR) var[k][d] = HMM_VAR_FLOOR;
}
}
#endif
}
};
class KMeansRegimeModel {
public:
double centroids[KMEANS_K][KMEANS_DIM];
double distEma;
double distVarEma;
int initialized;
int regime;
double dist;
double stability;
KMeansRegimeModel() : distEma(0), distVarEma(1), initialized(0), regime(0), dist(0), stability(0) {
memset(centroids, 0, sizeof(centroids));
}
void init() {
distEma = 0;
distVarEma = 1;
initialized = 0;
regime = 0;
dist = 0;
stability = 0;
memset(centroids, 0, sizeof(centroids));
}
void seed(const double x[KMEANS_DIM]) {
for(int k=0;k<KMEANS_K;k++) {
for(int d=0; d<KMEANS_DIM; d++) {
centroids[k][d] = x[d] + 0.03 * (k - 1);
}
}
initialized = 1;
}
static double clampRange(double x, double lo, double hi) {
if(x < lo) return lo;
if(x > hi) return hi;
return x;
}
void predictAndUpdate(const double x[KMEANS_DIM]) {
if(!initialized) seed(x);
int best = 0;
double bestDist = INF;
for(int k=0;k<KMEANS_K;k++) {
double s = 0;
for(int d=0; d<KMEANS_DIM; d++) {
double z = x[d] - centroids[k][d];
s += z * z;
}
double dk = sqrt(s + EPS);
if(dk < bestDist) {
bestDist = dk;
best = k;
}
}
regime = best;
dist = bestDist;
distEma = (1.0 - KMEANS_DIST_EMA) * distEma + KMEANS_DIST_EMA * dist;
double dd = dist - distEma;
distVarEma = (1.0 - KMEANS_DIST_EMA) * distVarEma + KMEANS_DIST_EMA * dd * dd;
double distStd = sqrt(distVarEma + EPS);
double zDist = (dist - distEma) / (distStd + EPS);
stability = clampRange(1.0 / (1.0 + exp(zDist)), 0.0, 1.0);
#if KMEANS_ONLINE_UPDATE
for(int d=0; d<KMEANS_DIM; d++) {
centroids[best][d] += KMEANS_ETA * (x[d] - centroids[best][d]);
}
#endif
}
};
class SpectralClusterModel {
public:
int clusterId[N_ASSETS];
int nClusters;
void init() {
nClusters = SPECTRAL_K;
for(int i=0;i<N_ASSETS;i++) clusterId[i] = i % SPECTRAL_K;
}
void update(const fvar* distMatrix) {
if(!distMatrix) return;
// lightweight deterministic clustering surrogate from distance rows
for(int i=0;i<N_ASSETS;i++) {
double sig = 0;
for(int j=0;j<N_ASSETS;j++) {
if(i == j) continue;
double d = (double)distMatrix[i*N_ASSETS + j];
if(d < INF) sig += d;
}
int cid = (int)fmod(fabs(sig * 1000.0), (double)SPECTRAL_K);
if(cid < 0) cid = 0;
if(cid >= SPECTRAL_K) cid = SPECTRAL_K - 1;
clusterId[i] = cid;
}
}
};
class StrategyController {
public:
UnsupervisedModel unsup;
RLAgent rl;
PCAModel pca;
GMMRegimeModel gmm;
HMMRegimeModel hmm;
KMeansRegimeModel kmeans;
int dynamicTopK;
double scoreScale;
int regime;
double adaptiveGamma;
double adaptiveAlpha;
double adaptiveBeta;
double adaptiveLambda;
double riskScale;
int cooldown;
StrategyController()
: dynamicTopK(TOP_K), scoreScale(1.0), regime(0),
adaptiveGamma(1.0), adaptiveAlpha(1.0), adaptiveBeta(1.0), adaptiveLambda(1.0), riskScale(1.0), cooldown(0) {}
static double clampRange(double x, double lo, double hi) {
if(x < lo) return lo;
if(x > hi) return hi;
return x;
}
void init() {
unsup.init();
rl.init();
pca.init();
gmm.init();
hmm.init();
kmeans.init();
dynamicTopK = TOP_K;
scoreScale = 1.0;
regime = 0;
adaptiveGamma = 1.0;
adaptiveAlpha = 1.0;
adaptiveBeta = 1.0;
adaptiveLambda = 1.0;
riskScale = 1.0;
cooldown = 0;
}
void buildGMMState(const LearningSnapshot& snap, int reg, double conf, double x[GMM_DIM]) {
x[0] = snap.meanScore;
x[1] = snap.meanCompactness;
x[2] = snap.meanVol;
x[3] = pca.dom;
x[4] = pca.rot;
x[5] = (double)reg / 2.0;
x[6] = conf;
x[7] = snap.meanScore - snap.meanCompactness;
}
void buildHMMObs(const LearningSnapshot& snap, int reg, double conf, double x[HMM_DIM]) {
x[0] = pca.latent[0];
x[1] = pca.latent[1];
x[2] = pca.latent[2];
x[3] = snap.meanVol;
x[4] = snap.meanScore;
x[5] = snap.meanCompactness;
x[6] = (double)reg / 2.0;
x[7] = conf;
}
void buildKMeansState(const LearningSnapshot& snap, int reg, double conf, double x[KMEANS_DIM]) {
x[0] = pca.latent[0];
x[1] = pca.latent[1];
x[2] = pca.latent[2];
x[3] = snap.meanVol;
x[4] = snap.meanScore;
x[5] = snap.meanCompactness;
x[6] = (double)reg / 2.0;
x[7] = conf;
}
void onUpdate(const LearningSnapshot& snap, fvar* scores, int nScores, int updateCount) {
#if USE_ML
double unsupConf = 0;
unsup.update(snap, ®ime, &unsupConf);
#if USE_PCA
pca.update(snap, regime, unsupConf);
#else
pca.dom = 0.5;
pca.rot = 0.0;
#endif
#if USE_GMM
double gx[GMM_DIM];
buildGMMState(snap, regime, unsupConf, gx);
gmm.infer(gx);
#if USE_HMM
double hx[HMM_DIM];
buildHMMObs(snap, regime, unsupConf, hx);
hmm.filter(hx);
#if USE_KMEANS
double kx[KMEANS_DIM];
buildKMeansState(snap, regime, unsupConf, kx);
kmeans.predictAndUpdate(kx);
#endif
#endif
// regime presets: [gamma, alpha, beta, lambda]
const double presets[GMM_K][4] = {
{1.05, 1.00, 0.95, 1.00},
{0.95, 1.05, 1.05, 0.95},
{1.00, 0.95, 1.10, 1.05}
};
adaptiveGamma = 0;
adaptiveAlpha = 0;
adaptiveBeta = 0;
adaptiveLambda = 0;
for(int k=0;k<GMM_K;k++) {
#if USE_HMM
adaptiveGamma += hmm.posterior[k] * presets[k][0];
adaptiveAlpha += hmm.posterior[k] * presets[k][1];
adaptiveBeta += hmm.posterior[k] * presets[k][2];
adaptiveLambda += hmm.posterior[k] * presets[k][3];
#else
adaptiveGamma += gmm.p[k] * presets[k][0];
adaptiveAlpha += gmm.p[k] * presets[k][1];
adaptiveBeta += gmm.p[k] * presets[k][2];
adaptiveLambda += gmm.p[k] * presets[k][3];
#endif
}
#if USE_HMM
double entNorm = hmm.entropy / log((double)HMM_K + EPS);
riskScale = clampRange(1.0 - 0.45 * entNorm, HMM_MIN_RISK, 1.0);
if(hmm.entropy > HMM_ENTROPY_TH || hmm.switchProb > HMM_SWITCH_TH) cooldown = HMM_COOLDOWN_UPDATES;
else if(cooldown > 0) cooldown--;
#else
double entNorm = gmm.entropy / log((double)GMM_K + EPS);
riskScale = clampRange(1.0 - GMM_ENTROPY_COEFF * entNorm, GMM_MIN_RISK, 1.0);
#endif
#else
adaptiveGamma = 1.0 + 0.35 * pca.dom - 0.25 * pca.rot;
adaptiveAlpha = 1.0 + 0.30 * pca.dom;
adaptiveBeta = 1.0 + 0.25 * pca.rot;
adaptiveLambda = 1.0 + 0.20 * pca.dom - 0.20 * pca.rot;
riskScale = 1.0;
#endif
adaptiveGamma = clampRange(adaptiveGamma, 0.80, 1.40);
adaptiveAlpha = clampRange(adaptiveAlpha, 0.85, 1.35);
adaptiveBeta = clampRange(adaptiveBeta, 0.85, 1.35);
adaptiveLambda = clampRange(adaptiveLambda, 0.85, 1.25);
#if USE_KMEANS
const double kmPreset[KMEANS_K][4] = {
{1.02, 1.00, 0.98, 1.00},
{1.08, 0.96, 0.95, 1.02},
{0.94, 1.08, 1.08, 0.92}
};
int kr = kmeans.regime;
if(kr < 0) kr = 0;
if(kr >= KMEANS_K) kr = KMEANS_K - 1;
double wkm = clampRange(kmeans.stability, 0.0, 1.0);
adaptiveGamma = (1.0 - wkm) * adaptiveGamma + wkm * kmPreset[kr][0];
adaptiveAlpha = (1.0 - wkm) * adaptiveAlpha + wkm * kmPreset[kr][1];
adaptiveBeta = (1.0 - wkm) * adaptiveBeta + wkm * kmPreset[kr][2];
adaptiveLambda = (1.0 - wkm) * adaptiveLambda + wkm * kmPreset[kr][3];
if(kmeans.stability < KMEANS_STABILITY_MIN) {
riskScale *= 0.85;
if(cooldown < 1) cooldown = 1;
}
#endif
rl.updateReward(snap.meanScore);
rl.lastAction = rl.chooseAction(updateCount);
int baseTopK = TOP_K;
if(rl.lastAction == 0) baseTopK = TOP_K - 2;
else if(rl.lastAction == 1) baseTopK = TOP_K;
else if(rl.lastAction == 2) baseTopK = TOP_K;
else baseTopK = TOP_K - 1;
double profileBias[5] = {1.00, 0.98, 0.99, 0.97, 1.02};
scoreScale = (1.0 + 0.06 * (adaptiveGamma - 1.0) + 0.04 * (adaptiveAlpha - 1.0) - 0.04 * (adaptiveBeta - 1.0))
* profileBias[STRATEGY_PROFILE] * riskScale;
if(pca.dom > 0.60) baseTopK -= 1;
if(pca.rot > 0.15) baseTopK -= 1;
#if USE_HMM
if(hmm.regime == 2) baseTopK -= 1;
if(cooldown > 0) baseTopK -= 1;
#if USE_KMEANS
if(kmeans.regime == 2) baseTopK -= 1;
#endif
#elif USE_GMM
if(gmm.bestRegime == 2) baseTopK -= 1;
#endif
dynamicTopK = baseTopK;
if(dynamicTopK < 1) dynamicTopK = 1;
if(dynamicTopK > TOP_K) dynamicTopK = TOP_K;
for(int i=0; i<nScores; i++) {
double s = (double)scores[i] * scoreScale;
if(s > 1.0) s = 1.0;
if(s < 0.0) s = 0.0;
scores[i] = (fvar)s;
}
#else
(void)snap; (void)scores; (void)nScores; (void)updateCount;
#endif
}
};
// ---------------------------- Strategy ----------------------------
class CrowdAverseStrategy {
public:
ExposureTable exposureTable;
FeatureBufferSoA featSoA;
OpenCLBackend openCL;
SlabAllocator<fvar> corrMatrix;
SlabAllocator<fvar> distMatrix;
SlabAllocator<fvar> compactness;
SlabAllocator<fvar> entropy;
SlabAllocator<fvar> scores;
SlabAllocator<float> featLinear;
SlabAllocator<float> corrLinear;
int barCount;
int updateCount;
StrategyController controller;
SpectralClusterModel spectral;
CrowdAverseStrategy() : barCount(0), updateCount(0) {}
void init() {
printf("CrowdAverse_v9: Initializing...\n");
exposureTable.init();
featSoA.init(N_ASSETS, FEAT_WINDOW);
corrMatrix.init(N_ASSETS * N_ASSETS);
distMatrix.init(N_ASSETS * N_ASSETS);
compactness.init(N_ASSETS);
entropy.init(N_ASSETS);
scores.init(N_ASSETS);
featLinear.init(FEAT_N * N_ASSETS * FEAT_WINDOW);
corrLinear.init(N_ASSETS * N_ASSETS);
openCL.init();
printf("CrowdAverse_v9: Ready (OpenCL=%d)\n", openCL.ready);
controller.init();
spectral.init();
barCount = 0;
updateCount = 0;
}
void shutdown() {
printf("CrowdAverse_v9: Shutting down...\n");
openCL.shutdown();
featSoA.shutdown();
corrMatrix.shutdown();
distMatrix.shutdown();
compactness.shutdown();
entropy.shutdown();
scores.shutdown();
featLinear.shutdown();
corrLinear.shutdown();
}
void computeFeatures(int assetIdx) {
asset((char*)ASSET_NAMES[assetIdx]);
vars C = series(priceClose(0));
vars V = series(Volatility(C, 20));
if(Bar < 50) return;
fvar r1 = (fvar)log(C[0] / C[1]);
fvar rN = (fvar)log(C[0] / C[12]);
fvar vol = (fvar)V[0];
fvar zscore = (fvar)((C[0] - C[50]) / (V[0] * 20.0 + EPS));
fvar rangeP = (fvar)((C[0] - C[50]) / (C[0] + EPS));
fvar flow = (fvar)(r1 * vol);
fvar regime = (fvar)((vol > 0.001) ? 1.0 : 0.0);
fvar volOfVol = (fvar)(vol * vol);
fvar persistence = (fvar)fabs(r1);
featSoA.push(0, assetIdx, r1);
featSoA.push(1, assetIdx, rN);
featSoA.push(2, assetIdx, vol);
featSoA.push(3, assetIdx, zscore);
featSoA.push(4, assetIdx, rangeP);
featSoA.push(5, assetIdx, flow);
featSoA.push(6, assetIdx, regime);
featSoA.push(7, assetIdx, volOfVol);
featSoA.push(8, assetIdx, persistence);
}
fvar computeEntropy(int assetIdx) {
fvar mean = 0;
for(int t=0; t<FEAT_WINDOW; t++) mean += featSoA.get(0, assetIdx, t);
mean /= FEAT_WINDOW;
fvar var = 0;
for(int t=0; t<FEAT_WINDOW; t++) { fvar d = featSoA.get(0, assetIdx, t) - mean; var += d*d; }
return (fvar)(var / FEAT_WINDOW);
}
void computeCorrelationMatrixCPU() {
for(int i=0;i<N_ASSETS*N_ASSETS;i++) corrMatrix[i] = 0;
for(int f=0; f<FEAT_N; f++){
for(int a=0; a<N_ASSETS; a++){
for(int b=a+1; b<N_ASSETS; b++){
fvar mx = 0, my = 0;
for(int t=0; t<FEAT_WINDOW; t++){
mx += featSoA.get(f,a,t);
my += featSoA.get(f,b,t);
}
mx /= (fvar)FEAT_WINDOW;
my /= (fvar)FEAT_WINDOW;
fvar sxx = 0, syy = 0, sxy = 0;
for(int t=0; t<FEAT_WINDOW; t++){
fvar dx = featSoA.get(f,a,t) - mx;
fvar dy = featSoA.get(f,b,t) - my;
sxx += dx*dx;
syy += dy*dy;
sxy += dx*dy;
}
fvar den = (fvar)sqrt((double)(sxx*syy + (fvar)EPS));
fvar corr = 0;
if(den > (fvar)EPS) corr = sxy / den;
else corr = 0;
int idx = a*N_ASSETS + b;
corrMatrix[idx] += corr / (fvar)FEAT_N;
corrMatrix[b*N_ASSETS + a] = corrMatrix[idx];
}
}
}
}
void buildFeatLinear() {
int idx = 0;
for(int f=0; f<FEAT_N; f++){
for(int a=0; a<N_ASSETS; a++){
for(int t=0; t<FEAT_WINDOW; t++){
featLinear[idx] = (float)featSoA.get(f, a, t);
idx++;
}
}
}
}
void computeCorrelationMatrix() {
if(openCL.ready) {
buildFeatLinear();
for(int i=0;i<N_ASSETS*N_ASSETS;i++) corrLinear[i] = 0.0f;
int ok = openCL.computeCorrelationMatrixCL(
featLinear.data,
corrLinear.data,
N_ASSETS,
FEAT_N,
FEAT_WINDOW
);
if(ok) {
for(int i=0;i<N_ASSETS*N_ASSETS;i++) corrMatrix[i] = (fvar)0;
for(int a=0; a<N_ASSETS; a++){
corrMatrix[a*N_ASSETS + a] = (fvar)1.0;
for(int b=a+1; b<N_ASSETS; b++){
float c = corrLinear[a*N_ASSETS + b];
corrMatrix[a*N_ASSETS + b] = (fvar)c;
corrMatrix[b*N_ASSETS + a] = (fvar)c;
}
}
return;
}
printf("OpenCL: runtime fail -> CPU fallback\n");
openCL.ready = 0;
}
computeCorrelationMatrixCPU();
}
void computeDistanceMatrix() {
for(int i=0;i<N_ASSETS;i++){
for(int j=0;j<N_ASSETS;j++){
if(i == j) {
distMatrix[i*N_ASSETS + j] = (fvar)0;
} else {
fvar corrDist = (fvar)1.0 - (fvar)fabs((double)corrMatrix[i*N_ASSETS + j]);
fvar expDist = (fvar)exposureTable.getDist(i, j);
fvar blended = (fvar)LAMBDA_META * corrDist + (fvar)(1.0 - (double)LAMBDA_META) * expDist;
distMatrix[i*N_ASSETS + j] = blended;
}
}
}
}
void floydWarshall() {
fvar d[28][28];
for(int i=0;i<N_ASSETS;i++){
for(int j=0;j<N_ASSETS;j++){
d[i][j] = distMatrix[i*N_ASSETS + j];
if(i == j) d[i][j] = (fvar)0;
if(d[i][j] < (fvar)0) d[i][j] = (fvar)INF;
}
}
for(int k=0;k<N_ASSETS;k++){
for(int i=0;i<N_ASSETS;i++){
for(int j=0;j<N_ASSETS;j++){
if(d[i][k] < (fvar)INF && d[k][j] < (fvar)INF) {
fvar nk = d[i][k] + d[k][j];
if(nk < d[i][j]) d[i][j] = nk;
}
}
}
}
for(int i=0;i<N_ASSETS;i++){
fvar w = 0;
for(int j=i+1;j<N_ASSETS;j++){
if(d[i][j] < (fvar)INF) w += d[i][j];
}
if(w > (fvar)0) compactness[i] = (fvar)(1.0 / (1.0 + (double)w));
else compactness[i] = (fvar)0;
entropy[i] = computeEntropy(i);
}
}
void computeScores() {
for(int i=0;i<N_ASSETS;i++){
fvar coupling = 0;
int count = 0;
for(int j=0;j<N_ASSETS;j++){
if(i != j && distMatrix[i*N_ASSETS + j] < (fvar)INF) {
coupling += compactness[j];
count++;
}
}
fvar pCouple = 0;
if(count > 0) pCouple = coupling / (fvar)count;
else pCouple = (fvar)0;
fvar C_A = compactness[i];
fvar Ent = entropy[i];
fvar rawScore = (fvar)ALPHA * Ent + (fvar)GAMMA * C_A - (fvar)BETA * pCouple;
if(rawScore > (fvar)30) rawScore = (fvar)30;
if(rawScore < (fvar)-30) rawScore = (fvar)-30;
scores[i] = (fvar)(1.0 / (1.0 + exp(-(double)rawScore)));
}
}
LearningSnapshot buildSnapshot() {
LearningSnapshot s;
s.meanScore = 0; s.meanCompactness = 0; s.meanVol = 0;
for(int i=0;i<N_ASSETS;i++) {
s.meanScore += (double)scores[i];
s.meanCompactness += (double)compactness[i];
s.meanVol += (double)featSoA.get(2, i, 0);
}
s.meanScore /= (double)N_ASSETS;
s.meanCompactness /= (double)N_ASSETS;
s.meanVol /= (double)N_ASSETS;
s.regime = 0;
s.regimeConfidence = 0;
return s;
}
void onBar() {
barCount++;
for(int i=0;i<N_ASSETS;i++) computeFeatures(i);
if(barCount % UPDATE_EVERY == 0) {
updateCount++;
computeCorrelationMatrix();
computeDistanceMatrix();
#if USE_SPECTRAL
spectral.update(distMatrix.data);
#endif
floydWarshall();
computeScores();
controller.onUpdate(buildSnapshot(), scores.data, N_ASSETS, updateCount);
printTopK();
}
}
void printTopK() {
int indices[N_ASSETS];
for(int i=0;i<N_ASSETS;i++) indices[i] = i;
int topN = controller.dynamicTopK;
for(int i=0;i<topN;i++){
for(int j=i+1;j<N_ASSETS;j++){
if(scores[indices[j]] > scores[indices[i]]) {
int tmp = indices[i];
indices[i] = indices[j];
indices[j] = tmp;
}
}
}
if(updateCount % 10 == 0) {
printf("===CrowdAverse_v9 Top-K(update#%d,OpenCL=%d)===\n",
updateCount, openCL.ready);
int selected[N_ASSETS];
int selCount = 0;
#if USE_SPECTRAL
int usedCluster[SPECTRAL_K];
for(int c=0;c<SPECTRAL_K;c++) usedCluster[c] = 0;
for(int i=0;i<topN;i++){
int idx = indices[i];
int cid = spectral.clusterId[idx];
if(cid < 0 || cid >= SPECTRAL_K) cid = 0;
if(usedCluster[cid]) continue;
usedCluster[cid] = 1;
selected[selCount++] = idx;
}
for(int i=0;i<topN && selCount<topN;i++){
int idx = indices[i];
int dup = 0;
for(int k=0;k<selCount;k++) if(selected[k]==idx){ dup=1; break; }
if(!dup) selected[selCount++] = idx;
}
#else
for(int i=0;i<topN;i++) selected[selCount++] = indices[i];
#endif
for(int i=0;i<selCount;i++){
int idx = selected[i];
printf(" %d.%s: score=%.4f, C=%.4f, Ent=%.6f\n", i+1, ASSET_NAMES[idx], (double)scores[idx], (double)compactness[idx], (double)entropy[idx]);
}
}
}
};
// ---------------------------- Zorro DLL entry ----------------------------
static CrowdAverseStrategy* S = NULL;
DLLFUNC void run()
{
if(is(INITRUN)) {
BarPeriod = 60;
LookBack = max(LookBack, FEAT_WINDOW + 50);
asset((char*)ASSET_NAMES[0]);
if(!S) {
S = new CrowdAverseStrategy();
S->init();
}
}
if(is(EXITRUN)) {
if(S) {
S->shutdown();
delete S;
S = NULL;
}
return;
}
if(!S || Bar < LookBack)
return;
S->onBar();
}
|
