Claim C19 · formal · entrylib:2026.000013@v3 · 2026-11-18

Corrected · 18 November 2026 at 13:08 UTC

Version 3 answers a reader's comments. Corrected: C3 no longer says sMAPE is minimised by the median for small errors, only near it, with a bound; C4 keeps the weighted median, and C12 now holds what was said about the ordinary median; the measure called msMAPE in version 2 is the Suilin loss. Added: how well a test set tells two forecasts apart as the actuals spread out (C18); the large-sample distribution of the sMAPE minimiser (C19); Smyl's msMAPE (C20) and what it rewards on intermittent demand (C21). Also added: sections, the definition of sMAPE first, and a limitation on C6.

By Ivan Svetunkov (ORCID 0000-0001-7826-0281)

What does sMAPE reward? An analytical and simulation study · C19

Principal claim C19 of this paper (Svetunkov, I. (guarantor), 2026), with what it rests on. Read the whole paper.

Text file for AI

Text file: the claim with what it rests on as plain text, to attach when your assistant cannot open the address.

The message

Read https://entrylib.org/p/2026.000013/C19, one claim of a research paper with the evidence it rests on. Tell me briefly what it says and how well it is supported. If you cannot open the link, tell me instead of guessing. Then wait for my questions.

The claim

SupportS 100C19 · formal · planned test · novelty: new, by the authors

Let ln⁡y\ln y have a density gg that is symmetric about ln⁡m\ln m, unimodal (g(ln⁡m+v1)≥g(ln⁡m+v2)g(\ln m + v_1) \ge g(\ln m + v_2) whenever |v1|≤|v2||v_1| \le |v_2|, strictly on a set of positive probability) and continuous at ln⁡m\ln m. Write x=ln⁡y−ln⁡mx = \ln y - \ln m, A=E[sech2⁡(x/2)tanh⁡(|x|/2)]A = \mathrm E[\operatorname{sech}^2(x/2)\tanh(|x|/2)] and B=E[sech4⁡(x/2)]B = \mathrm E[\operatorname{sech}^4(x/2)]. (a) A<2g(ln⁡m)A < 2g(\ln m). (b) Over test sets of nn independent actuals, the sMAPE minimiser of a test set f^n\hat f_n (a forecast with the lowest sMAPE of the test set) tends to mm in probability, and n(ln⁡f^n−ln⁡m)\sqrt n(\ln\hat f_n - \ln m) is asymptotically normal with mean 0 and variance B/(2g(ln⁡m)−A)2B/(2g(\ln m) - A)^2; for the sample median the variance is 1/(4g(ln⁡m)2)1/(4g(\ln m)^2). The ratio of these two limiting variances, the asymptotic relative efficiency of the sMAPE minimiser to the sample median, is ARE=(2g(ln⁡m)−A)24g(ln⁡m)2B;\mathrm{ARE} = \frac{(2g(\ln m) - A)^2}{4\,g(\ln m)^2\,B}; that the relative efficiency of finite test sets tends to it is not shown here, since the variances of finite test sets need not tend to those of the limits. (c) For lognormal actuals with standard deviation σ\sigma of ln⁡y\ln y: ARE→1\mathrm{ARE} \to 1 as σ→0\sigma \to 0 and AREσ3→6(ln⁡2)22π≈7.23\mathrm{ARE}\,\sigma^3 \to 6(\ln 2)^2\sqrt{2\pi} \approx 7.23 as σ→∞\sigma \to \infty. For log-Laplace actuals with scale bb of ln⁡y\ln y: ARE→1\mathrm{ARE} \to 1 as b→0b \to 0 and AREb→3\mathrm{ARE}\,b \to 3 as b→∞b \to \infty. So on both families the sMAPE minimiser is asymptotically as precise as the sample median for narrow distributions and asymptotically loses all of its precision relative to it as the actuals spread out.

SupportS 100C1 · formal · planned test · novelty: extends earlier work, by the authors

For y>0y > 0 and f>0f > 0, the sMAPE term s(y,f)=|y−f|/(|y|+|f|)s(y,f) = |y-f|/(|y|+|f|) depends on yy and ff only through their ratio r=y/fr = y/f: s=|r−1|/(r+1)s = |r - 1|/(r + 1).

SupportS 100C3 · formal · planned test · novelty: extends earlier work, by the authors

For y>0y > 0 and f>0f > 0, the sMAPE term is s=tanh⁡(|ln⁡y−ln⁡f|/2)s = \tanh(|\ln y - \ln f|/2) (log error ln⁡y−ln⁡f\ln y - \ln f). (a) It is 0 when y=fy = f, strictly increasing and concave in |ln⁡y−ln⁡f||\ln y - \ln f|, and bounded by 1, which it approaches as |ln⁡y−ln⁡f||\ln y - \ln f| grows. (b) 0≤12|ln⁡y−ln⁡f|−s≤124|ln⁡y−ln⁡f|30 \le \tfrac12|\ln y - \ln f| - s \le \tfrac1{24}|\ln y - \ln f|^3; so when all log errors are small, sMAPE is about 100 times the mean absolute error of ln⁡f\ln f as a forecast of ln⁡y\ln y (log-MAE). (c) The log-MAE of a test set is minimised by its median; sMAPE is then minimised near the median, not in general at it. For positive actuals y1,…,yny_1, \dots, y_n with a median mm, a forecast f*f^* that minimises their sMAPE satisfies 1n∑t|ln⁡yt−ln⁡f*|−1n∑t|ln⁡yt−ln⁡m|≤112⋅1n∑t|ln⁡yt−ln⁡f*|3:\frac1n\sum_t|\ln y_t - \ln f^*| - \frac1n\sum_t|\ln y_t - \ln m| \le \frac1{12}\cdot\frac1n\sum_t|\ln y_t - \ln f^*|^3: its mean absolute log error exceeds the smallest possible by at most a twelfth of its mean cubed absolute log error. The minimiser is the median itself only under further conditions on the distribution of the actuals.

SupportS 100C4 · formal · planned test · novelty: new, by the authors

Let the actual yy be positive and take no single value with positive probability. A forecast ff has the lowest expected sMAPE only if the actuals below ff and those above it carry equal total weight w(y,f)=y/(y+f)2w(y,f) = y/(y+f)^2: E[w(y,f)𝟏{y<f}]=E[w(y,f)𝟏{y>f}].\mathrm E[w(y,f)\,\mathbf 1\{y<f\}] = \mathrm E[w(y,f)\,\mathbf 1\{y>f\}]. So the sMAPE target is a median of the actuals weighted by ww, where the ordinary median gives every actual the same weight. An actual's weight is largest when it equals the forecast, is the same for y/f=ry/f = r and 1/r1/r, and falls as y/fy/f moves away from 1 in either direction.

SupportS 100C9 · formal · planned test · novelty: extends earlier work, by the authors

For actuals and forecasts of any sign, 0≤s≤10 \le s \le 1. So the expected sMAPE of every forecast is finite, between 0 and 200, whatever the distribution of the actuals and without any moment of it being finite (unlike RMSE, MAE and MAPE, C8). For a positive random actual yy, at least one forecast has the lowest expected sMAPE.

SupportS 100C12 · formal · planned test · novelty: new, by the authors

Let the actual yy be positive and take no single value with positive probability, let mm be its ordinary (unweighted) median, and let its expected sMAPE have a single local minimum. With the forecast at mm, compare the total weight w(y,m)=y/(y+m)2w(y,m) = y/(y+m)^2 of the actuals above mm with that of the actuals below it. (a) If the actuals above mm carry more weight, the sMAPE target lies above mm; if those below carry more, it lies below mm. (b) If the two weights are equal, mm is a point where expected sMAPE is flat. It is then the target if, in addition, 2g(ln⁡m)>E[sech2⁡(ln⁡y−ln⁡m2)tanh⁡(|ln⁡y−ln⁡m|2)],2g(\ln m) > \mathrm E\Big[\operatorname{sech}^2\Big(\frac{\ln y - \ln m}{2}\Big)\tanh\Big(\frac{|\ln y - \ln m|}{2}\Big)\Big], where gg is the density of ln⁡y\ln y, continuous at ln⁡m\ln m: this makes mm a strict local minimum. Without it, mm can be a flat point of inflection, with the target elsewhere. (c) When ln⁡y\ln y has a density symmetric about ln⁡m\ln m and unimodal, mm is the only target; lognormal and log-Laplace actuals are examples.

Claims it rests on

SupportS 100C5 · formal · planned test · novelty: replicates earlier work, by the authors

Let ln⁡y\ln y have a density gg symmetric about ln⁡m\ln m, g(ln⁡m+v)=g(ln⁡m−v)g(\ln m+v) = g(\ln m-v), and unimodal: g(ln⁡m+v1)≥g(ln⁡m+v2)g(\ln m+v_1) \ge g(\ln m+v_2) whenever |v1|≤|v2||v_1| \le |v_2|, strictly on a set of positive probability. Then the median mm of yy is the only sMAPE target.

Evidence

E1 · quant · other · cited by C1, C5, C9

Numerical checks of the formulas of the formal claims (formula checks), lab run RN8 under the formal plan that decided them; every check is within the plan's tolerances. They check the formulas on concrete numbers; the proofs are the evidence for the claims. The rows are labelled by the hypotheses of the record's first plan, whose wording kept the factor 200: H1 checks C1; H2 and H10 check C2; H5 checks C5; H7 checks C6; H8 checks C7; H12 checks C8; H13 checks C9; H14 checks C10. H3 and H9 checked C3, and H4 and H6 checked C4, as version 2 worded them; C3, C4 and C12 as now worded are checked in E31. The two results E1.R1 and E1.R2 compare a simulated variance with the large-sample variance that C19 now proves.

  • R1 simulated n Var(ln sMAPE minimiser) / large-sample value, lognormal sigma=1 = 0.9944 99% confidence interval 0.9354 to 1.0598 (hypothesis: H11, test-set size: 1000, test sets: 4000)
  • R2 simulated n Var(ln sMAPE minimiser) / large-sample value, lognormal sigma=1 = 1.022 99% confidence interval 0.9396 to 1.0982 (hypothesis: H11, test-set size: 10000, test sets: 2000)
  • R3 maximum absolute difference from the ratio form = 0.0000 (hypothesis: H1)
  • R4 maximum absolute difference between the ratio and its reciprocal = 0.0000 (hypothesis: H2)
  • R5 maximum absolute difference from 200 tanh(|z|/2) = 0.0000 (hypothesis: H3)
  • R6 violations of monotonicity, concavity or the bound 200 on 0 <= |z| <= 12 = 0 (hypothesis: H3)
  • R7 violations of 0 <= 100|z| - s <= 25|z|^3/3 = 0 (hypothesis: H9)
  • R8 maximum of |s - 100|z| + 25|z|^3/3| / |z|^5 for 0.05 < |z| < 0.5 = 0.8331 (hypothesis: H9)
  • R9 violations of s(y, y+e) < s(y, y-e) = 0 (hypothesis: H10)
  • R10 maximum absolute difference between s(y, ky) and s(y, y/k) = 0.0000 (hypothesis: H10)
  • R11 maximum relative difference between the formula for S' and a numerical derivative = 0.0000 (hypothesis: H4)
  • R12 maximum distance in ln f between the minimiser and the median = 0.0000 (hypothesis: H5)
  • R13 number of local minima, lognormal sigma=1 and log-Laplace b=1 = 1 (hypothesis: H5)
  • R14 maximum balance residual at the minimiser (pull below minus pull above) = 0.0000 (hypothesis: H6)
  • R15 settings with one local minimum where the sign of D misplaces the target = 0 (hypothesis: H6)
  • R16 maximum distance between the one-step and the exact quantile level = 0.0025 (hypothesis: H6)
  • R17 2 g(m) - E[sech^2(Z/2) tanh(|Z|/2)] for the equal-mode bimodal setting (negative: median a local maximum) = -0.3475 (hypothesis: H7)
  • R18 number of local minima of the equal-mode bimodal setting = 2 (hypothesis: H7)
  • R19 difference between the two minimal expected sMAPEs, equal-mode bimodal setting = 0.0000 (hypothesis: H7)
  • R20 maximum distance between the formula and the bisected threshold = 0.0000 (hypothesis: H8)
  • R21 opposite-sign or one-zero pairs not scoring 200 = 0 (hypothesis: H8)
  • R22 relative distance between the 1/y-weighted median and the minimiser of expected MAPE = 0.0001 (hypothesis: H12)
  • R23 difference of expected squared errors minus (f1-f2)(f1+f2-2E[y]) = 0 (hypothesis: H12)
  • R24 sMAPE terms outside [0, 200] = 0 (hypothesis: H13)
  • R25 expected sMAPE at f = 1e-6 and f = 1e10 (both near 200) = 200 (hypothesis: H13)
  • R26 the proof's bound on the error times sigma^3, 600 zeta(3) max|g''|, log-Gumbel = 309.9 (hypothesis: H14)
  • R27 limit of sigma^3 times the relative efficiency for normal log-outcomes, from the proof = 7.226 (hypothesis: H14)
E1.T1 · zero threshold by setting; hypothesis: H8 (lab runs RN8) (hypothesis: H8)
Rowsettingzero threshold
r1lognormal sigma=0.30.4687
r2lognormal sigma=0.50.4469
r3lognormal sigma=10.394
r4lognormal sigma=20.3075
E1.T2 · maximum |S - (200 - 800 ln2 h)| times sigma^3, log-Gumbel right, sigma^3 times the asymptotic relative efficiency, lognormal by sigma; hypothesis: H14 (lab runs RN8) (hypothesis: H14)
Rowsigmamaximum |S - (200 - 800 ln2 h)| times sigma^3, log-Gumbel rightsigma^3 times the asymptotic relative efficiency, lognormal
r15.0264.55.613
r210.0292.36.743
r320.0300.17.099
r440.0302.17.193

E31 · quant · other · cited by C3, C4, C12

Numerical checks of the formulas of the claims new or reworded in version 3 (formula checks), lab run RN13 (check_v3.py, seed 8) under the formal plan that decided C3, C4, C12, C17 and C20; every check is within the plan's tolerances. They check the formulas on concrete numbers; the proofs are the evidence for the claims. The tables are labelled by the hypotheses the run's script names: H3 (T1 and T2) checks C3; H4 (T3) checks C4; H34 (T4 and T5) checks C12; H35 (T6 and T7) checks C17; H42 (T8 and T9) checks C20. The tables of H39 (T10 to T12) and H40 (T13 to T17) check the formulas of C18 and C19, whose wording was decided after later checks (E32, E33).

E31.T1 · violations of the bound on the log-MAE of the sMAPE minimiser, largest excess of the minimiser's log-MAE over the median's, minus the bound; hypothesis: H3, test_sets: 2000 (lab runs RN13) (hypothesis: H3, test_sets: 2000)
Rowviolations of the bound on the log-MAE of the sMAPE minimiserlargest excess of the minimiser's log-MAE over the median's, minus the bound
r100.0000
E31.T2 · sum of sMAPE terms, forecast at the median, actuals 1, 1, e, e^4, e^4, sum of sMAPE terms, forecast at 1, actuals 1, 1, e, e^4, e^4; hypothesis: H3 (lab runs RN13) (hypothesis: H3)
Rowsum of sMAPE terms, forecast at the median, actuals 1, 1, e, e^4, e^4sum of sMAPE terms, forecast at 1, actuals 1, 1, e, e^4, e^4
r12.7352.39
E31.T3 · largest difference of the weights below and above the target, over their sum; hypothesis: H4, distributions: 6 (lab runs RN13) (hypothesis: H4, distributions: 6)
Rowlargest difference of the weights below and above the target, over their sum
r10.0000
E31.T4 · distributions with more than one local minimum, distributions where the target is not on the side of the median that carries more weight; hypothesis: H34, distributions: 6 (lab runs RN13) (hypothesis: H34, distributions: 6)
Rowdistributions with more than one local minimumdistributions where the target is not on the side of the median that carries more weight
r100
E31.T5 · largest distance in logs between the target and the median, lognormal and log-Laplace; hypothesis: H34, distributions: 16 (lab runs RN13) (hypothesis: H34, distributions: 16)
Rowlargest distance in logs between the target and the median, lognormal and log-Laplace
r10.0000
E31.T6 · Suilin term of a zero forecast of a zero actual; hypothesis: H35 (lab runs RN13) (hypothesis: H35)
RowSuilin term of a zero forecast of a zero actual
r10
E31.T7 · largest error of the identity of (a), largest error of the identity of (b), values of the Suilin term outside [0, 1); hypothesis: H35, pairs: 200000 (lab runs RN13) (hypothesis: H35, pairs: 200000)
Rowlargest error of the identity of (a)largest error of the identity of (b)values of the Suilin term outside [0, 1)
r10.000000
E31.T8 · largest error of the identity of (a), largest error of the identity of (b), values of the msMAPE term outside [0, 1], values of the msMAPE term not below 1 for actuals below nu, pairs with a count of at least 1 where the msMAPE and sMAPE terms differ by nu; hypothesis: H42, pairs: 200000 (lab runs RN13) (hypothesis: H42, pairs: 200000)
Rownulargest error of the identity of (a)largest error of the identity of (b)values of the msMAPE term outside [0, 1]values of the msMAPE term not below 1 for actuals below nupairs with a count of at least 1 where the msMAPE and sMAPE terms differ
r10.100000
r20.500000
r31.000000
E31.T9 · msMAPE term of a forecast of 1e-9 for a zero actual by nu; hypothesis: H42 (lab runs RN13) (hypothesis: H42)
RownumsMAPE term of a forecast of 1e-9 for a zero actual
r10.10.0000
r20.50.0000
r31.00.0000
E31.T10 · signal-to-noise ratio of sMAPE over its large-spread formula, signal-to-noise ratio of the absolute error over its limit by error_distribution, c1, c2; hypothesis: H39, sigma: 200 (lab runs RN13) (hypothesis: H39, sigma: 200)
Rowerror_distributionc1c2signal-to-noise ratio of sMAPE over its large-spread formulasignal-to-noise ratio of the absolute error over its limit
r1normal0.00.50.99990.9913
r2normal0.01.010.9956
r3normal-0.51.51.0010.9966
r4Gumbel0.00.50.99980.9647
r5Gumbel0.01.010.9937
r6Gumbel-0.51.510.9967
E31.T11 · relative spread of the absolute log error's signal-to-noise ratio over sigma = 2, 10, 50 by error_distribution, c1, c2; hypothesis: H39 (lab runs RN13) (hypothesis: H39)
Rowerror_distributionc1c2relative spread of the absolute log error's signal-to-noise ratio over sigma = 2, 10, 50
r1normal0.00.50.0000
r2normal0.01.00.0000
r3normal-0.51.50.0000
r4Gumbel0.00.50.0000
r5Gumbel0.01.00.0000
r6Gumbel-0.51.50.0000
E31.T12 · signal-to-noise ratio of the squared error over its formula, lognormal, exp(sigma^2 / 2) times the squared error's ratio, lognormal, formula by c1, c2, sigma; hypothesis: H39 (lab runs RN13) (hypothesis: H39)
Rowc1c2sigmasignal-to-noise ratio of the squared error over its formula, lognormalexp(sigma^2 / 2) times the squared error's ratio, lognormal, formula
r10.00.50.51
r20.01.00.51
r3-0.51.50.51
r40.00.51.01
r50.01.01.01
r6-0.51.51.01
r70.00.52.01
r80.01.02.01
r9-0.51.52.01
r100.01.04.00.9907
r110.01.06.01
E31.T13 · distributions compared with the asymptotic values recorded by RN7; hypothesis: H40 (lab runs RN13) (hypothesis: H40)
Rowdistributions compared with the asymptotic values recorded by RN7
r116
E31.T14 · distributions where A is not below 2 g at the median, largest difference from the asymptotic values recorded by RN7; hypothesis: H40, distributions: 16 (lab runs RN13) (hypothesis: H40, distributions: 16)
Rowdistributions where A is not below 2 g at the medianlargest difference from the asymptotic values recorded by RN7
r100.0001
E31.T15 · large-sample relative efficiency, formula by family, spread; hypothesis: H40 (lab runs RN13) (hypothesis: H40)
Rowfamilyspreadlarge-sample relative efficiency, formula
r1lognormal0.250.9696
r2lognormal0.50.8881
r3lognormal1.00.6601
r4lognormal1.50.454
r5lognormal2.00.3073
r6lognormal3.00.1474
r7lognormal4.00.0778
r8lognormal6.00.0279
r9log-Laplace0.250.9969
r10log-Laplace0.50.9765
r11log-Laplace1.00.8953
r12log-Laplace1.50.8036
r13log-Laplace2.00.7207
r14log-Laplace3.00.5903
r15log-Laplace4.00.4969
r16log-Laplace6.00.3755
E31.T16 · distance of the large-sample relative efficiency from 1 at spread 0.01 by family; hypothesis: H40 (lab runs RN13) (hypothesis: H40)
Rowfamilydistance of the large-sample relative efficiency from 1 at spread 0.01
r1lognormal0.0001
r2log-Laplace0.0000
E31.T17 · large-sample relative efficiency times sigma^3 over 6 (ln 2)^2 sqrt(2 pi), lognormal, large-sample relative efficiency times b over 3, log-Laplace; hypothesis: H40, spread: 200 (lab runs RN13) (hypothesis: H40, spread: 200)
Rowlarge-sample relative efficiency times sigma^3 over 6 (ln 2)^2 sqrt(2 pi), lognormallarge-sample relative efficiency times b over 3, log-Laplace
r10.99980.9906

E32 · quant · other

Numerical checks of the formulas of C18 (formula checks), lab run RN18 (check_h39.py) under the formal plan that decided C18; every check is within the plan's tolerances. They check the formulas on concrete numbers; the proof is the evidence for the claim. The run's script labels the checks H39, the record's first wording of C18; the formulas are those of C18.

E32.T1 · distance of sMAPE's signal-to-noise ratio over its large-spread formula from 1, distance of the absolute error's signal-to-noise ratio over its limit from 1 by error_distribution, c1, c2, sigma; hypothesis: H39 (lab runs RN18) (hypothesis: H39)
Rowerror_distributionc1c2sigmadistance of sMAPE's signal-to-noise ratio over its large-spread formula from 1distance of the absolute error's signal-to-noise ratio over its limit from 1
r1normal0.00.52000.00010.0087
r2normal0.00.520000.00000.0009
r3normal0.01.02000.00040.0044
r4normal0.01.020000.00000.0004
r5normal-0.51.52000.00120.0034
r6normal-0.51.520000.00010.0003
r7Gumbel0.00.52000.00020.0353
r8Gumbel0.00.520000.00000.0035
r9Gumbel0.01.02000.00020.0063
r10Gumbel0.01.020000.00000.0006
r11Gumbel-0.51.52000.00040.0033
r12Gumbel-0.51.520000.00000.0003
E32.T2 · relative spread of the absolute log error's signal-to-noise ratio over sigma = 2, 10, 50 by error_distribution, c1, c2; hypothesis: H39 (lab runs RN18) (hypothesis: H39)
Rowerror_distributionc1c2relative spread of the absolute log error's signal-to-noise ratio over sigma = 2, 10, 50
r1normal0.00.50.0000
r2normal0.01.00.0000
r3normal-0.51.50.0000
r4Gumbel0.00.50.0000
r5Gumbel0.01.00.0000
r6Gumbel-0.51.50.0000
E32.T3 · distance of the squared error's signal-to-noise ratio over its formula from 1, lognormal, distance of exp(sigma^2 / 2) times the squared error's ratio from 1, lognormal, formula by c1, c2, sigma; hypothesis: H39 (lab runs RN18) (hypothesis: H39)
Rowc1c2sigmadistance of the squared error's signal-to-noise ratio over its formula from 1, lognormaldistance of exp(sigma^2 / 2) times the squared error's ratio from 1, lognormal, formula
r10.00.50.50.0000
r20.01.00.50.0000
r3-0.51.50.50.0000
r40.00.51.00.0000
r50.01.01.00.0000
r6-0.51.51.00.0000
r70.00.52.00.0000
r80.01.02.00.0000
r9-0.51.52.00.0000
r100.01.04.00.0093
r110.01.06.00.0000
r120.01.08.00.0000

E33 · quant · other · cited by C19

Numerical checks of the formulas of C19 (formula checks), lab run RN16 (check_corrected.py, seed 8) under the formal plan that decided C19; every check is within the plan's tolerances. They check the formulas on concrete numbers; the proof is the evidence for the claim. The run's script labels the checks by the record's first wordings: H40 (T4 to T7 and the result R1) checks C19; H39 (T1 to T3) checks the formulas of C18, which the same plan tested in an earlier wording.

  • R1 distributions compared with the asymptotic values recorded by RN7 = 16 (hypothesis: H40)
E33.T1 · distance of sMAPE's signal-to-noise ratio over its large-spread formula from 1, distance of the absolute error's signal-to-noise ratio over its limit from 1 by error_distribution, c1, c2, sigma; hypothesis: H39 (lab runs RN16) (hypothesis: H39)
Rowerror_distributionc1c2sigmadistance of sMAPE's signal-to-noise ratio over its large-spread formula from 1distance of the absolute error's signal-to-noise ratio over its limit from 1
r1normal0.00.52000.00010.0087
r2normal0.00.520000.00000.0009
r3normal0.01.02000.00040.0044
r4normal0.01.020000.00000.0004
r5normal-0.51.52000.00120.0034
r6normal-0.51.520000.00010.0003
r7Gumbel0.00.52000.00020.0353
r8Gumbel0.00.520000.00000.0035
r9Gumbel0.01.02000.00020.0063
r10Gumbel0.01.020000.00000.0006
r11Gumbel-0.51.52000.00040.0033
r12Gumbel-0.51.520000.00000.0003
E33.T2 · relative spread of the absolute log error's signal-to-noise ratio over sigma = 2, 10, 50 by error_distribution, c1, c2; hypothesis: H39 (lab runs RN16) (hypothesis: H39)
Rowerror_distributionc1c2relative spread of the absolute log error's signal-to-noise ratio over sigma = 2, 10, 50
r1normal0.00.50.0000
r2normal0.01.00.0000
r3normal-0.51.50.0000
r4Gumbel0.00.50.0000
r5Gumbel0.01.00.0000
r6Gumbel-0.51.50.0000
E33.T3 · distance of the squared error's signal-to-noise ratio over its formula from 1, lognormal, distance of exp(sigma^2 / 2) times the squared error's ratio from 1, lognormal, formula by c1, c2, sigma; hypothesis: H39 (lab runs RN16) (hypothesis: H39)
Rowc1c2sigmadistance of the squared error's signal-to-noise ratio over its formula from 1, lognormaldistance of exp(sigma^2 / 2) times the squared error's ratio from 1, lognormal, formula
r10.00.50.50.0000
r20.01.00.50.0000
r3-0.51.50.50.0000
r40.00.51.00.0000
r50.01.01.00.0000
r6-0.51.51.00.0000
r70.00.52.00.0000
r80.01.02.00.0000
r9-0.51.52.00.0000
r100.01.04.00.0093
r110.01.06.00.0000
r120.01.08.00.0000
E33.T4 · large-sample relative efficiency, formula by family, spread; hypothesis: H40 (lab runs RN16) (hypothesis: H40)
Rowfamilyspreadlarge-sample relative efficiency, formula
r1lognormal0.250.9696
r2lognormal0.50.8881
r3lognormal1.00.6601
r4lognormal1.50.454
r5lognormal2.00.3073
r6lognormal3.00.1474
r7lognormal4.00.0778
r8lognormal6.00.0279
r9log-Laplace0.250.9969
r10log-Laplace0.50.9765
r11log-Laplace1.00.8953
r12log-Laplace1.50.8036
r13log-Laplace2.00.7207
r14log-Laplace3.00.5903
r15log-Laplace4.00.4969
r16log-Laplace6.00.3755
E33.T5 · distance of the large-sample relative efficiency from 1 at spread 0.01 by family; hypothesis: H40 (lab runs RN16) (hypothesis: H40)
Rowfamilydistance of the large-sample relative efficiency from 1 at spread 0.01
r1lognormal0.0001
r2log-Laplace0.0000
E33.T6 · distributions where A is not below 2 g at the median, largest difference from the asymptotic values recorded by RN7; hypothesis: H40, distributions: 16 (lab runs RN16) (hypothesis: H40, distributions: 16)
Rowdistributions where A is not below 2 g at the medianlargest difference from the asymptotic values recorded by RN7
r100.0001
E33.T7 · large-sample relative efficiency times sigma^3 over 6 (ln 2)^2 sqrt(2 pi), lognormal, large-sample relative efficiency times b over 3, log-Laplace; hypothesis: H40, spread: 200 (lab runs RN16) (hypothesis: H40, spread: 200)
Rowlarge-sample relative efficiency times sigma^3 over 6 (ln 2)^2 sqrt(2 pi), lognormallarge-sample relative efficiency times b over 3, log-Laplace
r10.99980.9906

E18 · formal · cited by C1

Proof of C1 (section H1 of the proof document A1, rendered in A2). Only the ratio of the actual to the forecast matters, not their size.

  1. Since yy and ff are positive, |y|+|f|=y+f|y| + |f| = y + f. So s=|y−f|/(y+f)s = |y - f|/(y + f). (algebra; from D1)
  2. Dividing the numerator and the denominator by f>0f > 0 gives |y−f|/f(y+f)/f\dfrac{|y-f|/f}{(y+f)/f}. Since f>0f > 0, |y−f|/f=|(y−f)/f|=|r−1||y - f|/f = |(y - f)/f| = |r - 1|. And (y+f)/f=r+1(y+f)/f = r + 1. (algebra; from S1)
  3. So s=|r−1|/(r+1)s = |r - 1|/(r + 1). The contribution depends only on the ratio of actual to forecast. (conclusion; from S2)

E20 · formal · cited by C3

Proof of C3 (section H3 of the proof document A1, rendered in A2). sMAPE measures the error on the log scale. For small errors it is half the absolute log error; for large errors it levels off.

  1. Write z=ln⁡y−ln⁡fz = \ln y - \ln f, so that r=y/f=ezr = y/f = e^z. By C1, s=|ez−1|ez+1s = \dfrac{|e^z - 1|}{e^z + 1}. (definition; from C1)
  2. Multiplying the numerator and the denominator of ez−1ez+1\dfrac{e^z - 1}{e^z + 1} by e−z/2e^{-z/2} gives ez/2−e−z/2ez/2+e−z/2=tanh⁡(z/2)\dfrac{e^{z/2} - e^{-z/2}}{e^{z/2} + e^{-z/2}} = \tanh(z/2). The denominator is positive, so s=|tanh⁡(z/2)|s = |\tanh(z/2)|. (algebra; from S1)
  3. tanh⁡\tanh changes only its sign when its argument changes sign: tanh⁡(−x)=−tanh⁡(x)\tanh(-x) = -\tanh(x), as the definition shows. So |tanh⁡(z/2)|=tanh⁡(|z|/2)|\tanh(z/2)| = \tanh(|z|/2), which is the stated form. (algebra; from S2)
  4. For (a): tanh⁡\tanh is 0 at 0. On [0,∞)[0, \infty) it is strictly increasing and concave, and it approaches 1 as its argument grows; its derivative 1−tanh2⁡x1 - \tanh^2 x is positive and decreasing for x>0x > 0. With the argument |ln⁡y−ln⁡f|/2|\ln y - \ln f|/2, ss has these properties in |ln⁡y−ln⁡f||\ln y - \ln f|. (monotonicity; from S3)
  5. For (b): for every x≥0x \ge 0, x−x3/3≤tanh⁡x≤xx - x^3/3 \le \tanh x \le x. The upper bound holds because tanh⁡\tanh has slope at most 1; the lower one because the derivative of tanh⁡x−x+x3/3\tanh x - x + x^3/3 is x2−tanh2⁡x≥0x^2 - \tanh^2 x \ge 0, and the difference is 0 at x=0x = 0. Putting x=|ln⁡y−ln⁡f|/2x = |\ln y - \ln f|/2 gives the bounds. (algebra; from S4)
  6. For (c), the median: the mean of |ln⁡yt−ln⁡f||\ln y_t - \ln f| over a test set is minimised when ln⁡f\ln f is a median of the ln⁡yt\ln y_t. Since the logarithm is increasing, this is when ff is a median mm of the yty_t. (monotonicity; from S5)
  7. For (c), the bound: with zt=ln⁡yt−ln⁡fz_t = \ln y_t - \ln f the log errors of the test set, write |z|‾(f)=1n∑t|ln⁡yt−ln⁡f|\overline{|z|}(f) = \frac1n\sum_t|\ln y_t - \ln f|, |z|3‾(f)=1n∑t|ln⁡yt−ln⁡f|3\overline{|z|^3}(f) = \frac1n\sum_t|\ln y_t - \ln f|^3 and s¯(f)\bar s(f) for the mean of the sMAPE terms of the test set. Averaging the bounds of step 5 over the test set gives s¯(f)≤12|z|‾(f)≤s¯(f)+124|z|3‾(f)\bar s(f) \le \tfrac12\overline{|z|}(f) \le \bar s(f) + \tfrac1{24}\overline{|z|^3}(f) for every f>0f > 0. Let f*f^* minimise s¯\bar s, so that s¯(f*)≤s¯(m)\bar s(f^*) \le \bar s(m). Then 12|z|‾(f*)≤s¯(f*)+124|z|3‾(f*)≤s¯(m)+124|z|3‾(f*)≤12|z|‾(m)+124|z|3‾(f*).\tfrac12\overline{|z|}(f^*) \le \bar s(f^*) + \tfrac1{24}\overline{|z|^3}(f^*) \le \bar s(m) + \tfrac1{24}\overline{|z|^3}(f^*) \le \tfrac12\overline{|z|}(m) + \tfrac1{24}\overline{|z|^3}(f^*). Multiplying by 2 gives |z|‾(f*)−|z|‾(m)≤112|z|3‾(f*)\overline{|z|}(f^*) - \overline{|z|}(m) \le \tfrac1{12}\overline{|z|^3}(f^*), the stated inequality. (algebra; from S5)
  8. For (c), not the median in general: take the five actuals 1,1,e,e4,e41, 1, e, e^4, e^4, whose median is ee. With the forecast at ee the log errors are 1,1,0,3,31, 1, 0, 3, 3 in size and the sMAPE terms sum to 2tanh⁡(0.5)+2tanh⁡(1.5)=2.7342\tanh(0.5) + 2\tanh(1.5) = 2.734. With the forecast at 1 the log errors are 0,0,1,4,40, 0, 1, 4, 4 and the terms sum to tanh⁡(0.5)+2tanh⁡(2)=2.390\tanh(0.5) + 2\tanh(2) = 2.390, which is lower. So the median does not minimise the sMAPE of this test set, although it minimises its log-MAE (step 6). (algebra; from S6)
  9. So s=tanh⁡(|ln⁡y−ln⁡f|/2)s = \tanh(|\ln y - \ln f|/2). It rises from 0 like half the absolute log error and levels off at 1. For small errors sMAPE is close to the log-MAE, and its minimiser is close to the median without in general being the median. (conclusion; from S8)

E21 · formal · cited by C9

Proof of C9 (section H4 of the proof document A1, rendered in A2). The expected values of RMSE, MAE and MAPE become infinite for every forecast when the actuals have an infinite moment, such as an infinite mean (a heavy tail, C8). The expected sMAPE never does, and there is always a best forecast.

  1. Finite. For positive values, s=tanh⁡(|ln⁡y−ln⁡f|/2)s = \tanh(|\ln y - \ln f|/2) (C3), which lies between 0 and 1. For any signs, |y−f|≤|y|+|f||y - f| \le |y| + |f|, so again s≤1s \le 1. A quantity between 0 and 1 has an expectation between 0 and 1, so the expected sMAPE is between 0 and 200, whatever the actuals. (earlier result; from C3)
  2. A best forecast. Write E[s(y,f)]\mathrm E[s(y,f)] for the expectation over the random actual yy. As ff approaches 0 or grows without bound, |ln⁡y−ln⁡f||\ln y - \ln f| grows for every positive value of yy, so ss approaches 1 (C3), and so does its expectation. For every positive ff it is below 1. It changes continuously with ff, because ss does and is at most 1 (bounded convergence). A continuous function that is below 1 in between and returns towards 1 at both ends reaches a lowest value at some f>0f > 0. A forecast of 0 or a negative forecast scores 1 on every positive actual, so it is never better. (bounded convergence; from C3, D1)
  3. So the expected sMAPE is always finite, whatever the tails of the actuals. For positive actuals, a forecast with the lowest expected sMAPE always exists. (conclusion; from S2)

E22 · formal · cited by C4

Proof of C4 (section H5 of the proof document A1, rendered in A2). For any distribution, the best forecast balances the actuals below it and above it. Each actual is weighted by how close it is to the forecast by ratio.

  1. The derivative of ss. When y>f>0y > f > 0, s=(y−f)/(y+f)s = (y - f)/(y + f). By the quotient rule its derivative in ff is −(y+f)−(y−f)(y+f)2=−2w(y,f)\dfrac{-(y+f) - (y-f)}{(y+f)^2} = -2w(y,f). When f>y>0f > y > 0, s=(f−y)/(y+f)s = (f - y)/(y + f), and in the same way its derivative is 2w(y,f)2w(y,f). Together: for y≠fy \ne f, the derivative is 2w(y,f)(𝟏{y<f}−𝟏{y>f})2w(y,f)(\mathbf 1\{y < f\} - \mathbf 1\{y > f\}). (quotient rule; from D1)
  2. A bound on it. Since (y+f)2≥4yf(y+f)^2 \ge 4yf, w(y,f)≤1/(4f)w(y,f) \le 1/(4f). So the derivative of ss is at most 1/(2f)1/(2f) in size, whatever yy, and at most 1/f1/f for every forecast above f/2f/2. The only actual at which ss has no derivative is y=fy = f, which has probability zero. (algebra; from S1)
  3. The derivative of the expectation. With this bound, the derivative may be taken inside the expectation (dominated convergence): ddfE[s(y,f)]=2E[w(y,f)(𝟏{y<f}−𝟏{y>f})].\frac{d}{df}\,\mathrm E[s(y,f)] = 2\,\mathrm E\big[w(y,f)\,(\mathbf 1\{y < f\} - \mathbf 1\{y > f\})\big]. (dominated convergence; from S2)
  4. The balance. A forecast with the lowest expected sMAPE exists (C9). There this derivative is 0, which is the stated balance: half the total weight lies below ff and half above. (earlier result; from C9)
  5. The weight. With r=y/fr = y/f, w(y,f)=1f⋅r(1+r)2=1f⋅1r+2+1/rw(y,f) = \dfrac1f\cdot\dfrac{r}{(1+r)^2} = \dfrac1f\cdot\dfrac{1}{r + 2 + 1/r}. The sum r+1/rr + 1/r is the same for rr and 1/r1/r. It is smallest, 2, at r=1r = 1, and grows as rr moves away from 1. So ww is largest at y=fy = f and falls as y/fy/f moves away from 1, equally for rr and 1/r1/r. (algebra; from S4)
  6. So sMAPE targets a weighted median of the actuals. Each actual is weighted by how close it is to the forecast by ratio. (conclusion; from S5)

E23 · formal · cited by C5

Proof of C5 (section H6 of the proof document A1, rendered in A2). This is the case of C4 in which the weights balance at the median.

  1. By C3, E[s(y,f)]=E[tanh⁡(|ln⁡y−ln⁡f|/2)]\mathrm E[s(y,f)] = \mathrm E\big[\tanh\big(|\ln y - \ln f|/2\big)\big]. Take a forecast f=meτf = m\,e^{\tau} with τ>0\tau > 0; the case τ<0\tau < 0 is the mirror image. We show that its expected sMAPE is larger than that of mm. (setup; from C3)
  2. Write both expectations as integrals over vv, the distance of ln⁡y\ln y from ln⁡f\ln f. For the forecast meτm\,e^{\tau}, ln⁡y=ln⁡m+τ+v\ln y = \ln m + \tau + v; for mm, ln⁡y=ln⁡m+v\ln y = \ln m + v. Their difference is Δ=∫tanh⁡(|v|/2)[g(ln⁡m+τ+v)−g(ln⁡m+v)]dv.\Delta = \int \tanh(|v|/2)\,\big[g(\ln m+\tau+v) - g(\ln m+v)\big]\,dv. (definition; from S1)
  3. Write ϕ(v)=g(ln⁡m+τ+v)−g(ln⁡m+v)\phi(v) = g(\ln m+\tau+v) - g(\ln m+v). By the symmetry of gg about ln⁡m\ln m, ϕ(−τ−v)=g(ln⁡m−v)−g(ln⁡m−τ−v)=g(ln⁡m+v)−g(ln⁡m+τ+v)=−ϕ(v)\phi(-\tau-v) = g(\ln m-v) - g(\ln m-\tau-v) = g(\ln m+v) - g(\ln m+\tau+v) = -\phi(v). So reflecting vv about −τ/2-\tau/2 reverses the sign of ϕ\phi. (definition; from S2)
  4. Split the integral at −τ/2-\tau/2 and reflect its left part onto the right. This gives Δ=∫v>−τ/2[tanh⁡(|v|/2)−tanh⁡(|v+τ|/2)]ϕ(v)dv.\Delta = \int_{v > -\tau/2} \big[\tanh(|v|/2) - \tanh(|v+\tau|/2)\big]\,\phi(v)\,dv. (algebra; from S3)
  5. For v>−τ/2v > -\tau/2, vv is closer to 0 than v+τv + \tau is: |v|<|v+τ||v| < |v + \tau|. So the bracket is negative, because tanh⁡\tanh is increasing (C3). And ϕ(v)≤0\phi(v) \le 0 by unimodality: g(ln⁡m+τ+v)≤g(ln⁡m+v)g(\ln m+\tau+v) \le g(\ln m+v). Their product is ≥0\ge 0, and >0> 0 on a set of positive probability. So Δ>0\Delta > 0. (monotonicity; from C3)
  6. So every forecast other than the median has a larger expected sMAPE. The median is the only forecast that minimises it. (conclusion; from S5)

E29 · formal · cited by C12

Proof of C12 (section H12 of the proof document A1, rendered in A2). The target is above the ordinary median when the actuals above it carry more weight, and below it when those below carry more. When the weights balance and enough actuals lie near the median, the median is the target. This is so when the log of the actual is symmetric with a single peak.

  1. The derivative at the median. By C4, step 3, the derivative of the expected sMAPE term in ff is 2E[w(y,f)(𝟏{y<f}−𝟏{y>f})]2\,\mathrm E[w(y,f)(\mathbf 1\{y < f\} - \mathbf 1\{y > f\})]. At f=mf = m it is twice the total weight of the actuals below mm minus that of the actuals above mm. (earlier result; from C4)
  2. The shape of expected sMAPE. With a single local minimum, the expected sMAPE decreases up to it and increases after it. Otherwise it would turn again, at a local maximum. After that turn it would have to fall and then rise once more, because it returns towards 1 at both ends (C9). That would give a second local minimum. (earlier result; from C9)
  3. (a). If the actuals above mm carry more weight, the derivative at the median is negative (step 1): the expected sMAPE is decreasing at the median. By step 2, the median then lies before the minimum, so the target is above the median. The case where the actuals below carry more weight is the mirror image. (symmetry; from S1, S2)
  4. The second derivative. Measure the forecast on the log scale, c=ln⁡fc = \ln f, so that the expected contribution is E[tanh⁡(|ln⁡y−c|/2)]\mathrm E[\tanh(|\ln y - c|/2)] (C3). The derivative of tanh⁡(|x−c|/2)\tanh(|x - c|/2) in cc is −12sech2⁡(x−c2)-\tfrac12\operatorname{sech}^2\big(\tfrac{x-c}{2}\big) for c<xc < x and +12sech2⁡(x−c2)+\tfrac12\operatorname{sech}^2\big(\tfrac{x-c}{2}\big) for c>xc > x: it jumps by 1 at c=xc = x. Differentiating again and taking the expectation over x=ln⁡yx = \ln y, the jumps add the density gg of ln⁡y\ln y at cc, and elsewhere the second derivative is −12sech2⁡(x−c2)tanh⁡(|x−c|2)-\tfrac12\operatorname{sech}^2\big(\tfrac{x-c}{2}\big)\tanh\big(\tfrac{|x-c|}{2}\big). So the second derivative of the expected contribution is g(c)−12E[sech2⁡(ln⁡y−c2)tanh⁡(|ln⁡y−c|2)].g(c) - \tfrac12\,\mathrm E\Big[\operatorname{sech}^2\Big(\frac{\ln y - c}{2}\Big)\tanh\Big(\frac{|\ln y - c|}{2}\Big)\Big]. This holds when gg is continuous at cc: since sech2⁡=1−tanh2⁡\operatorname{sech}^2 = 1 - \tanh^2, the first derivative is P(ln⁡y≤c)−12\mathrm P(\ln y \le c) - \tfrac12, whose derivative at cc is g(c)g(c), plus the expectation of −12tanh2⁡(x−c2)-\tfrac12\tanh^2\big(\tfrac{x-c}{2}\big) with the sign of c−xc - x, a function of cc with a bounded continuous derivative, which may be differentiated inside the expectation. (continuity; from C3)
  5. (b), the sufficient condition. If the two weights are equal, the first derivative at the median is 0 (step 1): the expected sMAPE is flat there. If, in addition, the stated inequality holds, the second derivative at c=ln⁡mc = \ln m is positive (step 4), so the median is a strict local minimum. With a single local minimum, it is the target. (algebra; from S1, S4)
  6. (b), without the inequality. A point where the first derivative is 0 can be a flat point of inflection, through which the expected sMAPE keeps rising (or keeps falling) without turning. An example, built by the independent checker of this section: let ln⁡y−ln⁡m\ln y - \ln m be a mixture of three normal distributions, with mean −1.5-1.5 and standard deviation 0.30.3 (weight 0.406125), mean 0 and standard deviation 0.50.5 (weight 0.054825), and mean 1.1586641.158664 and standard deviation 1 (weight 0.539050); these values are rounded from the exact solution of the three conditions. Its median is mm, the two weights are equal there, and the second derivative of step 4 is 0 at ln⁡m\ln m; the expected sMAPE rises through mm, where it is flat, and its only minimum is near ln⁡f=ln⁡m−1.25\ln f = \ln m - 1.25. So equal weights alone do not make the median the target. (algebra; from S4)
  7. (c). By C5 the median is then the only forecast with the lowest expected sMAPE (a density cannot be constant, so a symmetric unimodal density is strictly lower somewhere away from its centre, as C5 asks). For lognormal actuals ln⁡y\ln y has a normal density, and for log-Laplace actuals the Laplace density e−|x−ln⁡m|/b/(2b)e^{-|x - \ln m|/b}/(2b); both are symmetric about ln⁡m\ln m and strictly decreasing in the distance from it. (earlier result; from C5)
  8. So the target is above the ordinary median when the actuals above it carry more of the weight, and below it when those below carry more. When the two balance, the median is the target if, in addition, enough actuals lie near it, as the stated inequality says; this is the case when the log of the actual is symmetric with a single peak. (conclusion; from S7)

E36 · formal · cited by C19

Proof of C19 (section H14 of the proof document A1, rendered in A2). Where the log of the actual is symmetric with a single peak, the forecast that minimises the sMAPE of a test set is asymptotically normal about the median. Compared with the sample median it is asymptotically as precise for narrow distributions and asymptotically loses all of its precision as the actuals spread out.

  1. Write xt=ln⁡yt−ln⁡mx_t = \ln y_t - \ln m for the actuals of a test set and θ=ln⁡f−ln⁡m\theta = \ln f - \ln m for a forecast ff. By C3 the mean sMAPE term of the test set is Mn(θ)=1n∑ttanh⁡(|xt−θ|/2)M_n(\theta) = \frac1n\sum_t\tanh(|x_t - \theta|/2), and its expectation is M(θ)=E[tanh⁡(|x−θ|/2)]M(\theta) = \mathrm E[\tanh(|x - \theta|/2)]. So θ^n=ln⁡f^n−ln⁡m\hat\theta_n = \ln\hat f_n - \ln m minimises MnM_n; a minimiser exists, because MnM_n is continuous, tends to 1 as |θ||\theta| grows, and is below 1 at each xtx_t. (setup; from C3)
  2. (a). Write κ(u)=sech2⁡(u/2)tanh⁡(|u|/2)≥0\kappa(u) = \operatorname{sech}^2(u/2)\tanh(|u|/2) \ge 0. Since κ(u)\kappa(u) is the derivative of tanh2⁡(u/2)\tanh^2(u/2) for u>0u > 0, ∫κ(u)du=2\int\kappa(u)\,du = 2. By unimodality g(ln⁡m+u)≤g(ln⁡m)g(\ln m + u) \le g(\ln m), so A=∫κ(u)g(ln⁡m+u)du≤2g(ln⁡m)A = \int\kappa(u)\,g(\ln m + u)\,du \le 2g(\ln m). A density cannot equal g(ln⁡m)g(\ln m) everywhere, and g(ln⁡m+u)g(\ln m + u) does not increase as |u||u| grows; so g(ln⁡m+u)<g(ln⁡m)g(\ln m + u) < g(\ln m) for every |u||u| beyond some point, where κ>0\kappa > 0. So the inequality is strict. (definition; from S1)
  3. The expected loss is lowest at 0 only. M(θ)M(\theta) is the expected sMAPE term of the forecast meθm\,e^\theta. By C5, M(θ)>M(0)M(\theta) > M(0) for every θ≠0\theta \ne 0. The function tanh⁡(|u|/2)\tanh(|u|/2) has slope at most 12\tfrac12, so MM and every MnM_n change by at most 12|θ1−θ2|\tfrac12|\theta_1 - \theta_2| between θ1\theta_1 and θ2\theta_2; in particular MM is continuous. And M(θ)→1>M(0)M(\theta) \to 1 > M(0) as |θ|→∞|\theta| \to \infty, by bounded convergence. So for every η>0\eta > 0 the lowest value of MM over |θ|≥η|\theta| \ge \eta is above M(0)M(0). (bounded convergence; from C5)
  4. MnM_n is close to MM for all θ\theta at once. For each θ\theta, Mn(θ)→M(θ)M_n(\theta) \to M(\theta) with probability 1 (law of large numbers; the terms are bounded). Take a grid of spacing γ\gamma on [−R,R][-R, R]: by the slope bound of step 3, once MnM_n is within γ\gamma of MM at the grid points it is within 2γ2\gamma of MM on all of [−R,R][-R, R]. Outside [−R,R][-R, R]: an actual with |xt|≤R/2|x_t| \le R/2 has tanh⁡(|xt−θ|/2)≥tanh⁡(R/4)\tanh(|x_t - \theta|/2) \ge \tanh(R/4) when |θ|>R|\theta| > R, so Mn(θ)M_n(\theta) is at least tanh⁡(R/4)\tanh(R/4) times the share of such actuals, which tends to P(|x|≤R/2)\mathrm P(|x| \le R/2). Since M(0)<1M(0) < 1, for RR large this is above M(0)M(0) by a fixed margin. (case analysis; from S3)
  5. Consistency. Mn(θ^n)≤Mn(0)M_n(\hat\theta_n) \le M_n(0), which tends to M(0)M(0). By step 4, with probability tending to 1, |θ^n|≤R|\hat\theta_n| \le R and M(θ^n)M(\hat\theta_n) is within any given margin of M(0)M(0); by step 3, |θ^n|<η|\hat\theta_n| < \eta. So θ^n→0\hat\theta_n \to 0 in probability, which is f^n→m\hat f_n \to m. This is the argument of van der Vaart (1998, Theorem 5.7). (algebra; from S4, S3, B15)
  6. The curvature at 0. By C12, step 4, with c=ln⁡mc = \ln m, the second derivative of MM at 0 is g(ln⁡m)−12A=12(2g(ln⁡m)−A)g(\ln m) - \tfrac12 A = \tfrac12(2g(\ln m) - A), which is positive by step 2. The first derivative at 0 is 0, because 0 is the minimum (step 3). (earlier result; from S2, S3, C12)
  7. Asymptotic normality. For every x≠0x \ne 0, θ↦tanh⁡(|x−θ|/2)\theta \mapsto \tanh(|x - \theta|/2) has a derivative at θ=0\theta = 0, −ψ(x)-\psi(x) with ψ(x)=12sech2⁡(x/2)\psi(x) = \tfrac12\operatorname{sech}^2(x/2) for x>0x > 0 and −12sech2⁡(x/2)-\tfrac12\operatorname{sech}^2(x/2) for x<0x < 0; and x=0x = 0 has probability zero. For every xx it changes by at most 12|θ1−θ2|\tfrac12|\theta_1 - \theta_2| (step 3). MM has a second-order expansion at its minimum 0 with a positive second derivative (step 6). And θ^n\hat\theta_n minimises MnM_n and tends to 0 in probability (step 5). These are the conditions of the theorem on estimators that minimise a criterion which is Lipschitz in the parameter (van der Vaart, 1998, Theorem 5.23). It gives that nθ^n\sqrt n\,\hat\theta_n is asymptotically normal with mean 0 and variance E[ψ(x)2]\mathrm E[\psi(x)^2] divided by the square of the second derivative of MM at 0. (cited theorem; from S3, S6, S5, B15)
  8. The variance. E[ψ(x)2]=14E[sech4⁡(x/2)]=B/4\mathrm E[\psi(x)^2] = \tfrac14\mathrm E[\operatorname{sech}^4(x/2)] = B/4, and the square of the second derivative is 14(2g(ln⁡m)−A)2\tfrac14(2g(\ln m) - A)^2 (step 6). So the variance is B/(2g(ln⁡m)−A)2B/(2g(\ln m) - A)^2. (algebra; from S6)
  9. The sample median. The logarithm of the sample median of the yty_t is the sample median of the ln⁡yt\ln y_t, which minimises the mean of |xt−θ||x_t - \theta|. The same theorem applies with |x−θ||x - \theta| in place of tanh⁡(|x−θ|/2)\tanh(|x - \theta|/2): the derivative at 0 is −1-1 for x>0x > 0 and +1+1 for x<0x < 0, with square 1; the criterion changes by at most |θ1−θ2||\theta_1 - \theta_2|; and the second derivative of E|x−θ|\mathrm E|x - \theta| at 0 is 2g(ln⁡m)2g(\ln m), positive because gg is largest at ln⁡m\ln m. So the variance is 1/(4g(ln⁡m)2)1/(4g(\ln m)^2). (cited theorem; from S8, B15)
  10. (b), the ratio. Dividing the variance of step 9 by that of step 8 gives ARE=(2g(ln⁡m)−A)2/(4g(ln⁡m)2B)\mathrm{ARE} = (2g(\ln m) - A)^2/(4g(\ln m)^2 B). (algebra; from S9, S8)
  11. (c), narrow distributions. Write ln⁡y=ln⁡m+σε\ln y = \ln m + \sigma\varepsilon with ε\varepsilon standard normal, or Laplace with scale 1 and σ=b\sigma = b; then g(ln⁡m)=h(0)/σg(\ln m) = h(0)/\sigma with hh the density of ε\varepsilon, and x=σεx = \sigma\varepsilon. As σ→0\sigma \to 0, A→0A \to 0 and B→1B \to 1 by bounded convergence, and A/(2g(ln⁡m))=σA/(2h(0))→0A/(2g(\ln m)) = \sigma A/(2h(0)) \to 0. So ARE=(1−A/(2g(ln⁡m)))2/B→1\mathrm{ARE} = (1 - A/(2g(\ln m)))^2/B \to 1. (bounded convergence; from S10)
  12. (c), wide lognormal distributions. Since ∫κ=2\int\kappa = 2 (step 2), 2g(ln⁡m)−A=∫κ(u)[g(ln⁡m)−g(ln⁡m+u)]du2g(\ln m) - A = \int\kappa(u)\,[g(\ln m) - g(\ln m + u)]\,du. For the normal density, g(ln⁡m)−g(ln⁡m+u)=(1−e−u2/(2σ2))/(σ2π)g(\ln m) - g(\ln m + u) = \big(1 - e^{-u^2/(2\sigma^2)}\big)/(\sigma\sqrt{2\pi}), and u22σ2−u48σ4≤1−e−u2/(2σ2)≤u22σ2\tfrac{u^2}{2\sigma^2} - \tfrac{u^4}{8\sigma^4} \le 1 - e^{-u^2/(2\sigma^2)} \le \tfrac{u^2}{2\sigma^2}. Integrating by parts, ∫u2κ(u)du=4∫0∞usech2⁡(u/2)du=16ln⁡2\int u^2\kappa(u)\,du = 4\int_0^\infty u\operatorname{sech}^2(u/2)\,du = 16\ln 2; and ∫u4κ(u)du\int u^4\kappa(u)\,du is finite. So 2g(ln⁡m)−A=8ln⁡2/(σ32π)+O(σ−5)2g(\ln m) - A = 8\ln 2/(\sigma^3\sqrt{2\pi}) + O(\sigma^{-5}). In the same way, with ∫sech4⁡(u/2)du=8/3\int\operatorname{sech}^4(u/2)\,du = 8/3, B=8/(3σ2π)+O(σ−3)B = 8/(3\sigma\sqrt{2\pi}) + O(\sigma^{-3}). With g(ln⁡m)=1/(σ2π)g(\ln m) = 1/(\sigma\sqrt{2\pi}), AREσ3\mathrm{ARE}\,\sigma^3 tends to 64(ln⁡2)22π⋅2π4⋅32π8=6(ln⁡2)22π\dfrac{64(\ln 2)^2}{2\pi}\cdot\dfrac{2\pi}{4}\cdot\dfrac{3\sqrt{2\pi}}{8} = 6(\ln 2)^2\sqrt{2\pi}. (algebra; from S2)
  13. (c), wide log-Laplace distributions. Here g(ln⁡m+u)=e−|u|/b/(2b)g(\ln m + u) = e^{-|u|/b}/(2b), so g(ln⁡m)−g(ln⁡m+u)=(1−e−|u|/b)/(2b)g(\ln m) - g(\ln m + u) = (1 - e^{-|u|/b})/(2b), which lies between (|u|/b−u2/(2b2))/(2b)\big(|u|/b - u^2/(2b^2)\big)/(2b) and |u|/(2b2)|u|/(2b^2). Integrating by parts, ∫|u|κ(u)du=2∫0∞sech2⁡(u/2)du=4\int |u|\,\kappa(u)\,du = 2\int_0^\infty\operatorname{sech}^2(u/2)\,du = 4. So 2g(ln⁡m)−A=2/b2+O(b−3)2g(\ln m) - A = 2/b^2 + O(b^{-3}), and in the same way B=4/(3b)+O(b−2)B = 4/(3b) + O(b^{-2}). With g(ln⁡m)=1/(2b)g(\ln m) = 1/(2b), AREb\mathrm{ARE}\,b tends to 4b4⋅b2⋅3b4⋅b=3\dfrac{4}{b^4}\cdot b^2\cdot\dfrac{3b}{4}\cdot b = 3. (algebra; from S12)
  14. So the sMAPE minimiser of a test set is asymptotically normal about the median, with a variance that the density at the median and two expectations give. On lognormal and log-Laplace actuals it is asymptotically as precise as the sample median for narrow distributions and asymptotically loses all of its precision relative to it as the actuals spread out, at the rates σ−3\sigma^{-3} and b−1b^{-1}. (conclusion; from S13)

Definitions

Methods 1

D31 formula checks: Numerical checks of the formulas of the formal claims (claimset-code/smape-target/check_formulas.py, seed 6; check_jump.py for C11): identities and inequalities on 200000 random pairs, derivatives against finite differences, minimisers, balances and thresholds against quadrature and bisection on every setting of the selection design, and the large-spread expansion against its bound. For the claims new or reworded in version 3 (C3, C4, C12, C17 to C20): check_v3.py (seed 8), check_h39.py and check_corrected.py (seed 8), in the same way.

Variables and measures 13

D2 sMAPE: Symmetric mean absolute percentage error of a point forecast ff over a test set of actuals y1,…,yny_1, \dots, y_n: 200n∑t|yt−f|/(|yt|+|f|)\frac{200}{n}\sum_t |y_t - f|/(|y_t|+|f|), where a term with yt=f=0y_t = f = 0 counts as 0. Introduced as the symmetric MAPE by Makridakis (1993), whose denominator had no absolute values; for positive actuals and forecasts the two agree. Source: Makridakis (1993).

D1 sMAPE term: The contribution of one actual yy and one point forecast ff to sMAPE: s(y,f)=|y−f|/(|y|+|f|)s(y,f) = |y-f|/(|y|+|f|), between 0 and 1, with s=0s = 0 when y=fy = f (including y=f=0y = f = 0, where the fraction is 0/00/0).

D3 expected sMAPE: The expected sMAPE of a fixed forecast ff under the distribution of the actual yy: 200E[s(y,f)]200\,\mathrm{E}[s(y, f)], where E\mathrm E is the expectation over the random actual yy. Computed from the distribution (by integration or exact sums), never from a sample.

D4 log error: For a positive actual yy and a positive forecast ff, z=ln⁡y−ln⁡fz = \ln y - \ln f.

D5 weight: For a positive actual yy and a positive forecast ff, w(y,f)=y/(y+f)2w(y,f) = y/(y+f)^2: how much the actual pulls the forecast that minimises expected sMAPE, since the derivative of E[s(y,f)]\mathrm E[s(y,f)] in ff is 2E[w(y,f)(𝟏{y<f}−𝟏{y>f})]2\,\mathrm E[w(y,f)(\mathbf 1\{y<f\} - \mathbf 1\{y>f\})]. Largest at y=fy = f, the same for y/f=ry/f = r and 1/r1/r.

D10 RMSE: Root mean squared error of a point forecast ff over a test set of actuals y1,…,yny_1, \dots, y_n: 1n∑t(yt−f)2\sqrt{\frac1n\sum_t (y_t-f)^2}.

D11 MAE: Mean absolute error of a point forecast ff over a test set of actuals y1,…,yny_1, \dots, y_n: 1n∑t|yt−f|\frac1n\sum_t |y_t-f|.

D19 log-MAE: 100 times the mean absolute error of ln⁡f\ln f as a forecast of ln⁡y\ln y: 100n∑t|ln⁡yt−ln⁡f|\frac{100}{n}\sum_t |\ln y_t - \ln f|. Unbounded; sMAPE's small-error limit (C3); minimised by the median.

D28 Suilin term: The contribution of one actual yy and one forecast ff to the Suilin loss: s~(y,f)=|y−f|max⁡(|y|+|f|+δ,0.5+δ),δ=0.1;\tilde s(y,f) = \frac{|y-f|}{\max(|y|+|f|+\delta,\ 0.5+\delta)}, \qquad \delta = 0.1; the Suilin loss is 200 times its mean over a test set. It replaces the denominator |y|+|f||y|+|f| of the sMAPE term. Proposed by Suilin (2017) in code for a Kaggle competition, as a smoothed variant of sMAPE; defined in print by Hewamalage, Bergmeir and Bandara (2021, Eq. 15), who write the constant δ\delta as ϵ\epsilon and the factor 2 into the measure differently (a constant factor changes no property here). Hewamalage, Ackermann and Bergmeir (2023) call it msMAPE; Smyl (2025) gives that name to a different measure, so this study follows Kolassa (2026) and calls it the Suilin loss. Version 2 of this paper called it msMAPE. Source: Hewamalage et al. (2021).

D33 msMAPE term: The contribution of one actual y≥0y \ge 0 and one forecast f≥0f \ge 0 to the sparse-proof sMAPE (msMAPE) of Smyl (2025), as Kolassa (2026, Definition 4) states it, for a small constant ν>0\nu > 0 (Kolassa writes it mm, which is the median here): sν(y,f)=|f−y|f+max⁡{y,ν},s_\nu(y,f) = \frac{|f-y|}{f+\max\{y, \nu\}}, which is 0 when y=f=0y = f = 0; msMAPE is 200 times its mean over a test set (Kolassa writes the factor 2 into the term). It replaces the actual in the denominator of the sMAPE term by ν\nu when the actual is below ν\nu. Source: Kolassa (2026).

D8 sMAPE minimiser of a test set: The forecast that minimises the sMAPE of one test set y1,…,yny_1, \dots, y_n, f^n=arg⁡minf⁡1n∑ts(yt,f)\hat f_n = \arg\min_f \frac1n\sum_t s(y_t, f), found numerically (the sample counterpart of the sMAPE target).

D9 relative efficiency: For a test-set size nn and an outcome distribution where sMAPE and MAE both target the median: RE(n)=Var⁡(ln⁡m^n)/Var⁡(ln⁡f^n)\mathrm{RE}(n) = \operatorname{Var}(\ln \hat m_n) / \operatorname{Var}(\ln \hat f_n) over test sets, where m^n\hat m_n is the sample median (the MAE minimiser) and f^n\hat f_n the sMAPE minimiser of the same test set. Above 1, sMAPE's minimiser varies less; below 1, MAE's does.

D32 signal-to-noise ratio of a comparison: For a measure with term L(y,f)L(y,f), two forecasts f1f_1 and f2f_2 and a distribution of the actual yy: |E[D]|/sd(D)|\mathrm E[D]|/\mathrm{sd}(D), where D=L(y,f2)−L(y,f1)D = L(y,f_2) - L(y,f_1) is the difference of the two forecasts' scores on one actual. Over a test set of nn independent actuals the mean of DD is n\sqrt n times this many of its standard errors from zero: the larger the ratio, the smaller the test set that tells the two forecasts apart.

Concepts 1

D6 sMAPE target: The value or values of ff that minimise expected sMAPE for a given distribution of the actual yy, arg⁡minf⁡E[s(y,f)]\arg\min_f \mathrm E[s(y,f)]: the functional sMAPE rewards. It may be a single value or several.

Context

Extends B12 · from C1
Tofallis (2015) lists the sMAPE term, |y-f|/[(|y|+|f|)/2], among Tornqvist et al.'s measures of relative change; C1 shows it depends on the actual and the forecast only through their ratio.
Extends B12 · from C3
Tofallis (2015) relates sMAPE to the log accuracy ratio among relative measures; C3 shows the sMAPE term is exactly the hyperbolic tangent of half the absolute log ratio.
Explains B4 · from C4
C5 found sMAPE's preferred candidate close to the median on other distributions; C4 shows why: the target is a weighted median, on the side of the median that carries more weight (C12).
Addresses B12 · from C4
Tofallis (2015) states that minimising sMAPE does not produce an established measure of location; C4 identifies the target as a median weighted by y/(y+f)^2.
Qualifies B10 · from C4
Hewamalage et al. (2023) say sMAPE may tend towards selecting a slightly overestimating model; C4 shows the target is a weighted median, which lies on the side of the median that carries more weight, above or below it (C12).
Replicates B2 · from C5
C5 is an elementary case of Anderson's theorem on symmetric unimodal densities, proved here directly for the expected sMAPE.
Extends B10 · from C9
Hewamalage et al. (2023) note that sMAPE is bounded by 200; C9 adds that its expected value is therefore finite for every distribution and that a best forecast exists.
Explains B4 · from C12
C5 found by simulation that sMAPE prefers the median on lognormal and log-Laplace outcomes; C12 proves it for every symmetric unimodal ln y.

References

  1. Hewamalage, H., Bergmeir, C., Bandara, K., 2021. Recurrent Neural Networks for Time Series Forecasting: Current status and future directions. International Journal of Forecasting 37, 388–427. https://doi.org/10.1016/j.ijforecast.2020.06.008 Used by: D28
  2. Kolassa, S., 2026. The Symmetric Mean Absolute Percentage Error: Unnecessary or Dangerous. Forecasting 8, 24. https://doi.org/10.3390/forecast8020024 Used by: D33
  3. Makridakis, S., 1993. Accuracy measures: theoretical and practical concerns. International Journal of Forecasting 9, 527–529. https://doi.org/10.1016/0169-2070(93)90079-3 Used by: D2
  4. Vaart, A. W. v. d., 1998. Asymptotic Statistics. Cambridge University Press. https://doi.org/10.1017/CBO9780511802256 Used by: C19

Artefacts

  • A1 · proof Proofs of the formal claims, version 3, in Markdown with LaTeX: section H3 with part (c); H5 shortened to the weighted median; new sections H12 (the ordinary median, with the checker's example), H13 (signal-to-noise ratios) and H14 (large-sample efficiency). Checked independently by AC6 (CK27 to CK31, CK34 to CK37) and corrected as it proposed; after its checks the notation was made consistent (no second meaning for k, a, G), which AC6 read and found to change nothing. Sections H1 to H14 prove C1, C2, C3, C9, C4, C5, C6, C7, C11, C8, C10, C12, C18 and C19, in that order. Sections H1 to H11 were also checked independently by lab actor AC4 (proof-check-report.md) and by the supervisor. proofs.md
  • A2 · proof The same proofs, version 3, rendered as PDF. proofs.pdf
  • A6 · data Every result of the lab record (17182 results of 17 runs) at full precision: the lab result's ID, its run, the quantity as the run's script named it, its conditions and its value (both as JSON). The evidence's tables give the values to four significant digits. smape-target.results.csv

Score

Scores are computed by rules from the claimset file (EntryLib rubric 0.7): they check the record, not the science, and may change as the rubric and reviews develop. The rubric.

C19: 100, Well supported

Novelty: new, by the authors

Nothing lost.

C1: 100, Well supported

Novelty: extends earlier work, by the authors

Nothing lost.

C3: 100, Well supported

Novelty: extends earlier work, by the authors

Nothing lost.

C4: 100, Well supported

Novelty: new, by the authors

Nothing lost.

C5: 100, Well supported

Novelty: replicates earlier work, by the authors

Nothing lost.

C9: 100, Well supported

Novelty: extends earlier work, by the authors

Nothing lost.

C12: 100, Well supported

Novelty: new, by the authors

Nothing lost.