| 1 | // SPDX-License-Identifier: GPL-3.0-or-later |
| 2 | |
| 3 | #include "../libnetdata.h" |
| 4 | |
| 5 | NETDATA_DOUBLE default_single_exponential_smoothing_alpha = 0.1; |
| 6 | |
| 7 | void log_series_to_stderr(NETDATA_DOUBLE *series, size_t entries, NETDATA_DOUBLE result, const char *msg) { |
| 8 | const NETDATA_DOUBLE *value, *end = &series[entries]; |
| 9 | |
| 10 | fprintf(stderr, "%s of %zu entries [ ", msg, entries); |
| 11 | for(value = series; value < end ;value++) { |
| 12 | if(value != series) fprintf(stderr, ", "); |
| 13 | fprintf(stderr, "%" NETDATA_DOUBLE_MODIFIER, *value); |
| 14 | } |
| 15 | fprintf(stderr, " ] results in " NETDATA_DOUBLE_FORMAT "\n", result); |
| 16 | } |
| 17 | |
| 18 | // -------------------------------------------------------------------------------------------------------------------- |
| 19 | |
| 20 | inline NETDATA_DOUBLE sum_and_count(const NETDATA_DOUBLE *series, size_t entries, size_t *count) { |
| 21 | const NETDATA_DOUBLE *value, *end = &series[entries]; |
| 22 | NETDATA_DOUBLE sum = 0; |
| 23 | size_t c = 0; |
| 24 | |
| 25 | for(value = series; value < end ; value++) { |
| 26 | if(netdata_double_isnumber(*value)) { |
| 27 | sum += *value; |
| 28 | c++; |
| 29 | } |
| 30 | } |
| 31 | |
| 32 | if(unlikely(!c)) sum = NAN; |
| 33 | if(likely(count)) *count = c; |
| 34 | |
| 35 | return sum; |
| 36 | } |
| 37 | |
| 38 | inline NETDATA_DOUBLE sum(const NETDATA_DOUBLE *series, size_t entries) { |
| 39 | return sum_and_count(series, entries, NULL); |
| 40 | } |
| 41 | |
| 42 | inline NETDATA_DOUBLE average(const NETDATA_DOUBLE *series, size_t entries) { |
| 43 | size_t count = 0; |
| 44 | NETDATA_DOUBLE sum = sum_and_count(series, entries, &count); |
| 45 | |
| 46 | if(unlikely(!count)) return NAN; |
| 47 | return sum / (NETDATA_DOUBLE)count; |
| 48 | } |
| 49 | |
| 50 | // -------------------------------------------------------------------------------------------------------------------- |
| 51 | |
| 52 | // periods up to this size use a stack buffer to avoid heap overhead |
| 53 | #define MOVING_AVERAGE_STACK_PERIOD 32 |
| 54 | |
| 55 | NETDATA_DOUBLE moving_average(const NETDATA_DOUBLE *series, size_t entries, size_t period) { |
| 56 | // Keep the zero-period fast path while making the rolling window size |
| 57 | // unambiguously non-zero for the arithmetic below. |
| 58 | const size_t window = period ? period : 1; |
| 59 | |
| 60 | if(unlikely(period == 0)) |
| 61 | return 0.0; |
| 62 | |
| 63 | size_t i, count; |
| 64 | NETDATA_DOUBLE sum = 0, avg = 0; |
| 65 | |
| 66 | NETDATA_DOUBLE stack_buf[MOVING_AVERAGE_STACK_PERIOD]; |
| 67 | NETDATA_DOUBLE *heap_p = NULL; |
| 68 | NETDATA_DOUBLE *p; |
| 69 | |
| 70 | if(window <= MOVING_AVERAGE_STACK_PERIOD) { |
| 71 | memset(stack_buf, 0, window * sizeof(*stack_buf)); |
| 72 | p = stack_buf; |
| 73 | } else { |
| 74 | heap_p = callocz(window, sizeof(*heap_p)); |
| 75 | p = heap_p; |
| 76 | } |
| 77 | |
| 78 | for(i = 0, count = 0; i < entries; i++) { |
| 79 | NETDATA_DOUBLE value = series[i]; |
| 80 | if(unlikely(!netdata_double_isnumber(value))) continue; |
| 81 | size_t slot = count % window; |
| 82 | |
| 83 | if(unlikely(count < window)) { |
| 84 | sum += value; |
| 85 | avg = (count == window - 1) ? sum / (NETDATA_DOUBLE)window : 0; |
| 86 | } |
| 87 | else { |
| 88 | sum = sum - p[slot] + value; |
| 89 | avg = sum / (NETDATA_DOUBLE)window; |
| 90 | } |
| 91 | |
| 92 | p[slot] = value; |
| 93 | count++; |
| 94 | } |
| 95 | |
| 96 | freez(heap_p); |
| 97 | return avg; |
| 98 | } |
| 99 | |
| 100 | static int statistical_unittest_assert_close(const char *name, NETDATA_DOUBLE expected, NETDATA_DOUBLE actual) { |
| 101 | if(ABS(expected - actual) <= 0.000001) |
| 102 | return 0; |
| 103 | |
| 104 | fprintf(stderr, "statistical_unittest: %s failed, expected " NETDATA_DOUBLE_FORMAT ", got " NETDATA_DOUBLE_FORMAT "\n", |
| 105 | name, expected, actual); |
| 106 | return 1; |
| 107 | } |
| 108 | |
| 109 | int statistical_unittest(void) { |
| 110 | int errors = 0; |
| 111 | |
| 112 | NETDATA_DOUBLE series[] = { 1, 2, 3, 4, 5 }; |
| 113 | NETDATA_DOUBLE heap_series[MOVING_AVERAGE_STACK_PERIOD + 8]; |
| 114 | |
| 115 | for(size_t i = 0; i < sizeof(heap_series) / sizeof(heap_series[0]); i++) |
| 116 | heap_series[i] = (NETDATA_DOUBLE)(i + 1); |
| 117 | |
| 118 | errors += statistical_unittest_assert_close("moving_average(period=0)", 0.0, |
| 119 | moving_average(series, sizeof(series) / sizeof(series[0]), 0)); |
| 120 | errors += statistical_unittest_assert_close("moving_average(stack path)", 4.0, |
| 121 | moving_average(series, sizeof(series) / sizeof(series[0]), 3)); |
| 122 | errors += statistical_unittest_assert_close("moving_average(heap path)", |
| 123 | ((NETDATA_DOUBLE)(sizeof(heap_series) / sizeof(heap_series[0])) + 1.0) / 2.0, |
| 124 | moving_average(heap_series, sizeof(heap_series) / sizeof(heap_series[0]), |
| 125 | sizeof(heap_series) / sizeof(heap_series[0]))); |
| 126 | |
| 127 | if(errors) |
| 128 | fprintf(stderr, "statistical_unittest: %d errors found\n", errors); |
| 129 | else |
| 130 | fprintf(stderr, "statistical_unittest: all tests passed\n"); |
| 131 | |
| 132 | return errors ? 1 : 0; |
| 133 | } |
| 134 | |
| 135 | // -------------------------------------------------------------------------------------------------------------------- |
| 136 | |
| 137 | static int qsort_compare(const void *a, const void *b) { |
| 138 | NETDATA_DOUBLE *p1 = (NETDATA_DOUBLE *)a, *p2 = (NETDATA_DOUBLE *)b; |
| 139 | NETDATA_DOUBLE n1 = *p1, n2 = *p2; |
| 140 | |
| 141 | if(unlikely(isnan(n1) || isnan(n2))) { |
| 142 | if(isnan(n1) && !isnan(n2)) return -1; |
| 143 | if(!isnan(n1) && isnan(n2)) return 1; |
| 144 | return 0; |
| 145 | } |
| 146 | if(unlikely(isinf(n1) || isinf(n2))) { |
| 147 | if(!isinf(n1) && isinf(n2)) return -1; |
| 148 | if(isinf(n1) && !isinf(n2)) return 1; |
| 149 | return 0; |
| 150 | } |
| 151 | |
| 152 | if(unlikely(n1 < n2)) return -1; |
| 153 | if(unlikely(n1 > n2)) return 1; |
| 154 | return 0; |
| 155 | } |
| 156 | |
| 157 | inline void sort_series(NETDATA_DOUBLE *series, size_t entries) { |
| 158 | qsort(series, entries, sizeof(NETDATA_DOUBLE), qsort_compare); |
| 159 | } |
| 160 | |
| 161 | inline NETDATA_DOUBLE *copy_series(const NETDATA_DOUBLE *series, size_t entries) { |
| 162 | NETDATA_DOUBLE *copy = mallocz(sizeof(NETDATA_DOUBLE) * entries); |
| 163 | memcpy(copy, series, sizeof(NETDATA_DOUBLE) * entries); |
| 164 | return copy; |
| 165 | } |
| 166 | |
| 167 | NETDATA_DOUBLE percentile_on_sorted_series(const NETDATA_DOUBLE *series, size_t entries, double percentile) { |
| 168 | if (unlikely(entries == 0)) return NAN; |
| 169 | if (unlikely(entries == 1)) return series[0]; |
| 170 | |
| 171 | // Clamp percentile between 0.0 and 1.0 |
| 172 | percentile = fmax(0.0, fmin(1.0, percentile)); |
| 173 | |
| 174 | // Compute fractional index |
| 175 | NETDATA_DOUBLE index = percentile * (NETDATA_DOUBLE)(entries - 1); |
| 176 | size_t low_idx = (size_t)floor(index); |
| 177 | size_t high_idx = (size_t)ceil(index);; |
| 178 | |
| 179 | // If index is an integer or at the last element, return directly |
| 180 | if (high_idx >= entries || low_idx == high_idx || considered_equal_ndd(index, (NETDATA_DOUBLE)low_idx)) |
| 181 | return series[low_idx]; |
| 182 | |
| 183 | // Linear interpolation |
| 184 | NETDATA_DOUBLE weight = index - (NETDATA_DOUBLE)low_idx; |
| 185 | return series[low_idx] + weight * (series[high_idx] - series[low_idx]); |
| 186 | } |
| 187 | |
| 188 | NETDATA_DOUBLE median_on_sorted_series(const NETDATA_DOUBLE *series, size_t entries) { |
| 189 | return percentile_on_sorted_series(series, entries, 0.5); |
| 190 | } |
| 191 | |
| 192 | NETDATA_DOUBLE median(const NETDATA_DOUBLE *series, size_t entries) { |
| 193 | if(unlikely(entries == 0)) return NAN; |
| 194 | if(unlikely(entries == 1)) return series[0]; |
| 195 | |
| 196 | if(unlikely(entries == 2)) |
| 197 | return (series[0] + series[1]) / 2; |
| 198 | |
| 199 | NETDATA_DOUBLE *copy = copy_series(series, entries); |
| 200 | sort_series(copy, entries); |
| 201 | |
| 202 | NETDATA_DOUBLE avg = median_on_sorted_series(copy, entries); |
| 203 | |
| 204 | freez(copy); |
| 205 | return avg; |
| 206 | } |
| 207 | |
| 208 | // -------------------------------------------------------------------------------------------------------------------- |
| 209 | |
| 210 | NETDATA_DOUBLE moving_median(const NETDATA_DOUBLE *series, size_t entries, size_t period) { |
| 211 | if(entries <= period) |
| 212 | return median(series, entries); |
| 213 | |
| 214 | NETDATA_DOUBLE *data = copy_series(series, entries); |
| 215 | |
| 216 | size_t i; |
| 217 | for(i = period; i < entries; i++) { |
| 218 | data[i - period] = median(&series[i - period], period); |
| 219 | } |
| 220 | |
| 221 | NETDATA_DOUBLE avg = median(data, entries - period); |
| 222 | freez(data); |
| 223 | return avg; |
| 224 | } |
| 225 | |
| 226 | // -------------------------------------------------------------------------------------------------------------------- |
| 227 | |
| 228 | // http://stackoverflow.com/a/15150143/4525767 |
| 229 | NETDATA_DOUBLE running_median_estimate(const NETDATA_DOUBLE *series, size_t entries) { |
| 230 | NETDATA_DOUBLE median = 0.0f; |
| 231 | NETDATA_DOUBLE average = 0.0f; |
| 232 | size_t i; |
| 233 | |
| 234 | for(i = 0; i < entries ; i++) { |
| 235 | NETDATA_DOUBLE value = series[i]; |
| 236 | if(unlikely(!netdata_double_isnumber(value))) continue; |
| 237 | |
| 238 | average += ( value - average ) * 0.1f; // rough running average. |
| 239 | median += copysignndd( average * 0.01, value - median ); |
| 240 | } |
| 241 | |
| 242 | return median; |
| 243 | } |
| 244 | |
| 245 | // -------------------------------------------------------------------------------------------------------------------- |
| 246 | |
| 247 | NETDATA_DOUBLE standard_deviation(const NETDATA_DOUBLE *series, size_t entries) { |
| 248 | if(unlikely(entries == 0)) return NAN; |
| 249 | if(unlikely(entries == 1)) return series[0]; |
| 250 | |
| 251 | const NETDATA_DOUBLE *value, *end = &series[entries]; |
| 252 | size_t count; |
| 253 | NETDATA_DOUBLE sum; |
| 254 | |
| 255 | for(count = 0, sum = 0, value = series ; value < end ;value++) { |
| 256 | if(likely(netdata_double_isnumber(*value))) { |
| 257 | count++; |
| 258 | sum += *value; |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | if(unlikely(count == 0)) return NAN; |
| 263 | if(unlikely(count == 1)) return sum; |
| 264 | |
| 265 | NETDATA_DOUBLE average = sum / (NETDATA_DOUBLE)count; |
| 266 | |
| 267 | for(count = 0, sum = 0, value = series ; value < end ;value++) { |
| 268 | if(netdata_double_isnumber(*value)) { |
| 269 | count++; |
| 270 | sum += powndd(*value - average, 2); |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | if(unlikely(count == 0)) return NAN; |
| 275 | if(unlikely(count == 1)) return average; |
| 276 | |
| 277 | NETDATA_DOUBLE variance = sum / (NETDATA_DOUBLE)(count); // remove -1 from count to have a population stddev |
| 278 | NETDATA_DOUBLE stddev = sqrtndd(variance); |
| 279 | return stddev; |
| 280 | } |
| 281 | |
| 282 | // -------------------------------------------------------------------------------------------------------------------- |
| 283 | |
| 284 | NETDATA_DOUBLE single_exponential_smoothing(const NETDATA_DOUBLE *series, size_t entries, NETDATA_DOUBLE alpha) { |
| 285 | if(unlikely(entries == 0)) |
| 286 | return NAN; |
| 287 | |
| 288 | if(unlikely(isnan(alpha))) |
| 289 | alpha = default_single_exponential_smoothing_alpha; |
| 290 | |
| 291 | const NETDATA_DOUBLE *value = series, *end = &series[entries]; |
| 292 | NETDATA_DOUBLE level = (1.0 - alpha) * (*value); |
| 293 | |
| 294 | for(value++ ; value < end; value++) { |
| 295 | if(likely(netdata_double_isnumber(*value))) |
| 296 | level = alpha * (*value) + (1.0 - alpha) * level; |
| 297 | } |
| 298 | |
| 299 | return level; |
| 300 | } |
| 301 | |
| 302 | NETDATA_DOUBLE single_exponential_smoothing_reverse(const NETDATA_DOUBLE *series, size_t entries, NETDATA_DOUBLE alpha) { |
| 303 | if(unlikely(entries == 0)) |
| 304 | return NAN; |
| 305 | |
| 306 | if(unlikely(isnan(alpha))) |
| 307 | alpha = default_single_exponential_smoothing_alpha; |
| 308 | |
| 309 | const NETDATA_DOUBLE *value = &series[entries -1]; |
| 310 | NETDATA_DOUBLE level = (1.0 - alpha) * (*value); |
| 311 | |
| 312 | for(value++ ; value >= series; value--) { |
| 313 | if(likely(netdata_double_isnumber(*value))) |
| 314 | level = alpha * (*value) + (1.0 - alpha) * level; |
| 315 | } |
| 316 | |
| 317 | return level; |
| 318 | } |
| 319 | |
| 320 | // -------------------------------------------------------------------------------------------------------------------- |
| 321 | |
| 322 | // http://grisha.org/blog/2016/02/16/triple-exponential-smoothing-forecasting-part-ii/ |
| 323 | NETDATA_DOUBLE double_exponential_smoothing(const NETDATA_DOUBLE *series, size_t entries, |
| 324 | NETDATA_DOUBLE alpha, |
| 325 | NETDATA_DOUBLE beta, |
| 326 | NETDATA_DOUBLE *forecast) { |
| 327 | if(unlikely(entries == 0)) |
| 328 | return NAN; |
| 329 | |
| 330 | NETDATA_DOUBLE level, trend; |
| 331 | |
| 332 | if(unlikely(isnan(alpha))) |
| 333 | alpha = 0.3; |
| 334 | |
| 335 | if(unlikely(isnan(beta))) |
| 336 | beta = 0.05; |
| 337 | |
| 338 | level = series[0]; |
| 339 | |
| 340 | if(likely(entries > 1)) |
| 341 | trend = series[1] - series[0]; |
| 342 | else |
| 343 | trend = 0; |
| 344 | |
| 345 | const NETDATA_DOUBLE *value = series; |
| 346 | for(value++ ; value >= series; value--) { |
| 347 | if(likely(netdata_double_isnumber(*value))) { |
| 348 | NETDATA_DOUBLE last_level = level; |
| 349 | level = alpha * *value + (1.0 - alpha) * (level + trend); |
| 350 | trend = beta * (level - last_level) + (1.0 - beta) * trend; |
| 351 | |
| 352 | } |
| 353 | } |
| 354 | |
| 355 | if(forecast) |
| 356 | *forecast = level + trend; |
| 357 | |
| 358 | return level; |
| 359 | } |
| 360 | |
| 361 | // -------------------------------------------------------------------------------------------------------------------- |
| 362 | |
| 363 | /* |
| 364 | * Based on th R implementation |
| 365 | * |
| 366 | * a: level component |
| 367 | * b: trend component |
| 368 | * s: seasonal component |
| 369 | * |
| 370 | * Additive: |
| 371 | * |
| 372 | * Yhat[t+h] = a[t] + h * b[t] + s[t + 1 + (h - 1) mod p], |
| 373 | * a[t] = α (Y[t] - s[t-p]) + (1-α) (a[t-1] + b[t-1]) |
| 374 | * b[t] = β (a[t] - a[t-1]) + (1-β) b[t-1] |
| 375 | * s[t] = γ (Y[t] - a[t]) + (1-γ) s[t-p] |
| 376 | * |
| 377 | * Multiplicative: |
| 378 | * |
| 379 | * Yhat[t+h] = (a[t] + h * b[t]) * s[t + 1 + (h - 1) mod p], |
| 380 | * a[t] = α (Y[t] / s[t-p]) + (1-α) (a[t-1] + b[t-1]) |
| 381 | * b[t] = β (a[t] - a[t-1]) + (1-β) b[t-1] |
| 382 | * s[t] = γ (Y[t] / a[t]) + (1-γ) s[t-p] |
| 383 | */ |
| 384 | static int __HoltWinters( |
| 385 | const NETDATA_DOUBLE *series, |
| 386 | int entries, // start_time + h |
| 387 | |
| 388 | NETDATA_DOUBLE alpha, // alpha parameter of Holt-Winters Filter. |
| 389 | NETDATA_DOUBLE |
| 390 | beta, // beta parameter of Holt-Winters Filter. If set to 0, the function will do exponential smoothing. |
| 391 | NETDATA_DOUBLE |
| 392 | gamma, // gamma parameter used for the seasonal component. If set to 0, an non-seasonal model is fitted. |
| 393 | |
| 394 | const int *seasonal, |
| 395 | const int *period, |
| 396 | const NETDATA_DOUBLE *a, // Start value for level (a[0]). |
| 397 | const NETDATA_DOUBLE *b, // Start value for trend (b[0]). |
| 398 | NETDATA_DOUBLE *s, // Vector of start values for the seasonal component (s_1[0] ... s_p[0]) |
| 399 | |
| 400 | /* return values */ |
| 401 | NETDATA_DOUBLE *SSE, // The final sum of squared errors achieved in optimizing |
| 402 | NETDATA_DOUBLE *level, // Estimated values for the level component (size entries - t + 2) |
| 403 | NETDATA_DOUBLE *trend, // Estimated values for the trend component (size entries - t + 2) |
| 404 | NETDATA_DOUBLE *season // Estimated values for the seasonal component (size entries - t + 2) |
| 405 | ) |
| 406 | { |
| 407 | if(unlikely(entries < 4)) |
| 408 | return 0; |
| 409 | |
| 410 | int start_time = 2; |
| 411 | |
| 412 | NETDATA_DOUBLE res = 0, xhat = 0, stmp = 0; |
| 413 | int i, i0, s0; |
| 414 | |
| 415 | /* copy start values to the beginning of the vectors */ |
| 416 | level[0] = *a; |
| 417 | if(beta > 0) trend[0] = *b; |
| 418 | if(gamma > 0) memcpy(season, s, *period * sizeof(NETDATA_DOUBLE)); |
| 419 | |
| 420 | for(i = start_time - 1; i < entries; i++) { |
| 421 | /* indices for period i */ |
| 422 | i0 = i - start_time + 2; |
| 423 | s0 = i0 + *period - 1; |
| 424 | |
| 425 | /* forecast *for* period i */ |
| 426 | xhat = level[i0 - 1] + (beta > 0 ? trend[i0 - 1] : 0); |
| 427 | stmp = gamma > 0 ? season[s0 - *period] : (*seasonal != 1); |
| 428 | if (*seasonal == 1) |
| 429 | xhat += stmp; |
| 430 | else |
| 431 | xhat *= stmp; |
| 432 | |
| 433 | /* Sum of Squared Errors */ |
| 434 | res = series[i] - xhat; |
| 435 | *SSE += res * res; |
| 436 | |
| 437 | /* estimate of level *in* period i */ |
| 438 | if (*seasonal == 1) |
| 439 | level[i0] = alpha * (series[i] - stmp) |
| 440 | + (1 - alpha) * (level[i0 - 1] + trend[i0 - 1]); |
| 441 | else |
| 442 | level[i0] = alpha * (series[i] / stmp) |
| 443 | + (1 - alpha) * (level[i0 - 1] + trend[i0 - 1]); |
| 444 | |
| 445 | /* estimate of trend *in* period i */ |
| 446 | if (beta > 0) |
| 447 | trend[i0] = beta * (level[i0] - level[i0 - 1]) |
| 448 | + (1 - beta) * trend[i0 - 1]; |
| 449 | |
| 450 | /* estimate of seasonal component *in* period i */ |
| 451 | if (gamma > 0) { |
| 452 | if (*seasonal == 1) |
| 453 | season[s0] = gamma * (series[i] - level[i0]) |
| 454 | + (1 - gamma) * stmp; |
| 455 | else |
| 456 | season[s0] = gamma * (series[i] / level[i0]) |
| 457 | + (1 - gamma) * stmp; |
| 458 | } |
| 459 | } |
| 460 | |
| 461 | return 1; |
| 462 | } |
| 463 | |
| 464 | NETDATA_DOUBLE holtwinters(const NETDATA_DOUBLE *series, size_t entries, |
| 465 | NETDATA_DOUBLE alpha, |
| 466 | NETDATA_DOUBLE beta, |
| 467 | NETDATA_DOUBLE gamma, |
| 468 | NETDATA_DOUBLE *forecast) { |
| 469 | if(unlikely(isnan(alpha))) |
| 470 | alpha = 0.3; |
| 471 | |
| 472 | if(unlikely(isnan(beta))) |
| 473 | beta = 0.05; |
| 474 | |
| 475 | if(unlikely(isnan(gamma))) |
| 476 | gamma = 0; |
| 477 | |
| 478 | int seasonal = 0; |
| 479 | int period = 0; |
| 480 | NETDATA_DOUBLE a0 = series[0]; |
| 481 | NETDATA_DOUBLE b0 = 0; |
| 482 | NETDATA_DOUBLE s[] = {}; |
| 483 | |
| 484 | NETDATA_DOUBLE errors = 0.0; |
| 485 | size_t nb_computations = entries; |
| 486 | NETDATA_DOUBLE *estimated_level = callocz(nb_computations, sizeof(NETDATA_DOUBLE)); |
| 487 | NETDATA_DOUBLE *estimated_trend = callocz(nb_computations, sizeof(NETDATA_DOUBLE)); |
| 488 | NETDATA_DOUBLE *estimated_season = callocz(nb_computations, sizeof(NETDATA_DOUBLE)); |
| 489 | |
| 490 | int ret = __HoltWinters( |
| 491 | series, |
| 492 | (int)entries, |
| 493 | alpha, |
| 494 | beta, |
| 495 | gamma, |
| 496 | &seasonal, |
| 497 | &period, |
| 498 | &a0, |
| 499 | &b0, |
| 500 | s, |
| 501 | &errors, |
| 502 | estimated_level, |
| 503 | estimated_trend, |
| 504 | estimated_season |
| 505 | ); |
| 506 | |
| 507 | NETDATA_DOUBLE value = estimated_level[nb_computations - 1]; |
| 508 | |
| 509 | if(forecast) |
| 510 | *forecast = 0.0; |
| 511 | |
| 512 | freez(estimated_level); |
| 513 | freez(estimated_trend); |
| 514 | freez(estimated_season); |
| 515 | |
| 516 | if(!ret) |
| 517 | return 0.0; |
| 518 | |
| 519 | return value; |
| 520 | } |