Rolling projection calibration

Find the projection setup whose median is closest to actual outcomes.

Replay completed anchors and rank horizon × sample-size settings by projection error.

Snapshot 2026-10-01

Calibrated as-of 2026-10-01

Window 2026-07-24 → 2026-09-28

Anchors 120

Lookback 730D

Confidence low

Computed at 2026年10月2日上午7:47 HKT

Point-in-time Point-in-time safe

Best rolling setting

3D · n=40

Calibration results snapshot · Window: 2026-07-24 → 2026-09-28 · anchors: 120 · latest snapshot: 2026-10-01 · computed at: 2026年10月2日上午7:47 HKT

Median abs error2.55%

Mean abs error3.75%

Bias-0.19%

Inside P25/P7545.8%

Calibration setting comparison

Same idea as score-vs-outcome analysis: compare each parameter setting by observed actual-minus-median error over rolling completed anchors. Lower median abs error and lower bias win.

Setting Median abs Bias P25/P75 hit n Read

Setting 3D · n=40

Median abs 2.55%

Bias -0.19%

P25/P75 hit 45.8%

n 120

Rank Best setting

Setting 3D · n=80

Median abs 2.64%

Bias -0.14%

P25/P75 hit 45%

n 120

Rank #2 Candidate

Setting 3D · n=120

Median abs 2.78%

Bias -0.10%

P25/P75 hit 45.8%

n 120

Rank #3 Candidate

Setting 3D · n=320

Median abs 2.85%

Bias -0.22%

P25/P75 hit 41.7%

n 120

Rank #4 Candidate

Setting 3D · n=160

Median abs 2.89%

Bias -0.15%

P25/P75 hit 43.3%

n 120

Rank #5 Candidate

Setting 3D · n=240

Median abs 2.86%

Bias -0.20%

P25/P75 hit 42.5%

n 120

Rank #6 Candidate

Setting 3D · n=500

Median abs 2.87%

Bias -0.25%

P25/P75 hit 40%

n 120

Rank #7 Candidate

Setting 5D · n=320

Median abs 2.8%

Bias -0.24%

P25/P75 hit 50.9%

n 112

Rank #8 Candidate

Setting 5D · n=240

Median abs 2.81%

Bias -0.17%

P25/P75 hit 50%

n 112

Rank #9 Candidate

Setting 5D · n=40

Median abs 2.66%

Bias -0.05%

P25/P75 hit 43.8%

n 112

Rank #10 Candidate

Setting 5D · n=500

Median abs 2.86%

Bias -0.36%

P25/P75 hit 50%

n 112

Rank #11 Candidate

Setting 5D · n=80

Median abs 2.97%

Bias -0.04%

P25/P75 hit 48.2%

n 112

Rank #12 Candidate

Best setting 3D · n=40: median projection vs actual return

X = Median projection return. Y = actual return. Solid line = linear trend. Orange line = bucket average. Checked: 120 · Pending: 0 · Avg: -0.96%

Best setting 3D · n=40: median projection vs forecast error

X = Median projection return. Y = actual - median. Solid line = linear trend. Orange line = bucket average. Checked: 120 · Pending: 0 · Avg: -0.19%

Calibration ranking

Ranked by a composite Calibration score: lower median/mean absolute error, lower bias, enough anchors, and better IQR coverage.

Rank	Horizon	n	Median abs	Mean abs	Bias	Inside P25/P75	Anchors	Calibration score
#1	3D	40	2.55%	3.75%	-0.19%	45.8%	120	2.502
#2	3D	80	2.64%	3.71%	-0.14%	45%	120	2.527
#3	3D	120	2.78%	3.72%	-0.10%	45.8%	120	2.593
#4	3D	320	2.85%	3.71%	-0.22%	41.7%	120	2.652
#5	3D	160	2.89%	3.73%	-0.15%	43.3%	120	2.665
#6	3D	240	2.86%	3.75%	-0.20%	42.5%	120	2.666
#7	3D	500	2.87%	3.7%	-0.25%	40%	120	2.668
#8	5D	320	2.8%	4.27%	-0.24%	50.9%	112	2.803
#9	5D	240	2.81%	4.3%	-0.17%	50%	112	2.811
#10	5D	40	2.66%	4.53%	-0.05%	43.8%	112	2.813
#11	5D	500	2.86%	4.24%	-0.36%	50%	112	2.843
#12	5D	80	2.97%	4.39%	-0.04%	48.2%	112	2.907
#13	5D	160	3.03%	4.33%	-0.11%	50.9%	112	2.92
#14	5D	120	3.07%	4.35%	-0.08%	50%	112	2.944
#15	10D	240	3.61%	5.06%	+0.03%	54.2%	96	3.445
#16	10D	320	3.62%	5.03%	-0.10%	54.2%	96	3.45
#17	10D	500	3.69%	4.94%	-0.30%	55.2%	96	3.481
#18	10D	160	3.85%	5.2%	+0.09%	54.2%	96	3.623
#19	10D	120	3.96%	5.16%	+0.10%	56.3%	96	3.66
#20	10D	40	3.88%	5.45%	+0.46%	51%	96	3.789
#21	10D	80	4.09%	5.26%	+0.29%	51%	96	3.802
#22	15D	500	4.77%	6.04%	-1.28%	51.3%	80	4.563
#23	15D	320	4.94%	6.13%	-0.83%	52.5%	80	4.609
#24	15D	120	5.14%	6.51%	-0.15%	55%	80	4.733
#25	15D	240	5.29%	6.28%	-0.63%	52.5%	80	4.806
#26	15D	40	5%	6.87%	+0.24%	50%	80	4.816
#27	15D	160	5.27%	6.46%	-0.32%	51.3%	80	4.816
#28	15D	80	5.44%	6.68%	0.00%	56.3%	80	4.917
#29	20D	160	5.56%	6.72%	-0.47%	58%	69	5.058
#30	20D	120	5.57%	6.8%	-0.27%	55.1%	69	5.068