Learning · as of 2026-10-04
Backtest: 35 weekly origins (2026-01-05 → 2026-08-31), 28-day horizon, workload (volume × handle time) on every gate-channel-day. The cycle never reads the planted truth.
What the evidence says
- The record is hard to beat on a quiet estate. Workload WAPE is 10.46% for the record and 10.48% with every learned correction applied everywhere (-0.02 pts). Most gates have nothing real to learn, so applying corrections everywhere chases noise.
- Learning has to be earned, gate by gate. Under the walk-forward policy (a gate uses its corrections only after they beat the record in earlier windows) the estate scores 10.45% (+0.01 pts). That is the policy the proposal now uses, together with any confirmed shift in the last 70 days. 1 of 60 gate is learning today.
- The explicit overlays carry the big gains. Migration ratios, calendar and client signals are modeled changes with a source. They are not inferred from noise.
- Best layer in the backtest: earned (walk-forward) (10.45%).
Backtest by layer (workload WAPE, %)
| Scope | record | + learned volume | + learned AHT | + migration ratio | earned (walk-forward) | shift-gated |
|---|---|---|---|---|---|---|
| All gates | 10.46 | 10.48 | 10.48 | 10.48 | 10.45 | 10.46 |
| Product A | 11.52 | 11.56 | 11.56 | 11.56 | 11.53 | 11.52 |
| Product C | 11.15 | 11.17 | 11.18 | 11.18 | 11.15 | 11.15 |
| Product D | 8.99 | 9.02 | 9.03 | 9.03 | 8.99 | 8.99 |
| Product E | 10.01 | 10.00 | 9.99 | 9.99 | 9.96 | 10.01 |
| Bias, all gates | -1.07 | -1.55 | -1.60 | -1.60 | -1.13 | -1.03 |
Forecast value added (FVA) = WAPE of the record − WAPE of the layer, in points; positive means the layer helped.
Forward check
No actuals after the as-of date yet. Run with --actuals-through yyyy-mm-dd once they arrive.
Auto-approval: earned by 51 of 60 gates
A gate may be approved without review when its proposal moves less than 1% and the proposal beat or matched the record in at least 80% of its last 12 windows, never losing one by more than 2 points.
| Gate | Product | Windows | Beat or matched | Worst loss | Record WAPE | Proposal WAPE | Status |
|---|---|---|---|---|---|---|---|
| G102 | C | 12 | 12 | 0.4 pts | 13.47% | 13.57% | earned |
| G104 | D | 12 | 12 | 0.1 pts | 10.52% | 10.52% | earned |
| G105 | A | 12 | 12 | 0.24 pts | 8.55% | 8.59% | earned |
| G106 | A | 12 | 12 | 0.3 pts | 14.29% | 14.36% | earned |
| G107 | A | 12 | 12 | 0.22 pts | 13.2% | 13.28% | earned |
| G108 | A | 12 | 12 | 0.25 pts | 11.14% | 11.26% | earned |
| G109 | A | 12 | 12 | 0.32 pts | 17.32% | 17.38% | earned |
| G112 | A | 12 | 12 | 0.07 pts | 11.11% | 11.11% | earned |
| G113 | A | 12 | 12 | 0.5 pts | 14.59% | 14.62% | earned |
| G115 | A | 12 | 12 | 0.41 pts | 11.35% | 11.29% | earned |
| G116 | A | 12 | 12 | 0.37 pts | 13.54% | 13.45% | earned |
| G117 | A | 12 | 12 | 0.11 pts | 12.15% | 12.14% | earned |
| G118 | A | 12 | 12 | 0.16 pts | 11.53% | 11.53% | earned |
| G119 | A | 12 | 10 | 0.77 pts | 15.43% | 15.5% | earned |
| G120 | A | 12 | 12 | 0.01 pts | 12.26% | 12.14% | earned |
| G122 | A | 12 | 11 | 0.54 pts | 16.6% | 16.69% | earned |
| G123 | A | 12 | 11 | 0.6 pts | 12.82% | 12.96% | earned |
| G124 | A | 12 | 12 | 0.33 pts | 12.11% | 12.31% | earned |
| G125 | A | 12 | 12 | 0.33 pts | 19.64% | 19.69% | earned |
| G126 | A | 12 | 12 | 0.19 pts | 10.11% | 10.18% | earned |
| G128 | A | 12 | 12 | 0.25 pts | 13.78% | 13.75% | earned |
| G129 | A | 12 | 12 | 0.35 pts | 11.67% | 11.71% | earned |
| G130 | A | 12 | 12 | 0.18 pts | 10.28% | 10.31% | earned |
| G131 | A | 12 | 11 | 0.93 pts | 15.4% | 15.33% | earned |
| G132 | A | 12 | 12 | 0.19 pts | 10.52% | 10.38% | earned |
| G133 | A | 12 | 12 | 0.23 pts | 13.1% | 13.13% | earned |
| G134 | A | 12 | 12 | 0.28 pts | 15.11% | 15.17% | earned |
| G135 | A | 12 | 12 | 0.15 pts | 15.48% | 15.43% | earned |
| G136 | C | 12 | 11 | 0.64 pts | 10.32% | 10.45% | earned |
| G137 | C | 12 | 12 | 0.33 pts | 13.16% | 13.23% | earned |
| G138 | C | 12 | 12 | 0.35 pts | 12.01% | 12.01% | earned |
| G139 | C | 12 | 11 | 0.51 pts | 15.18% | 15.4% | earned |
| G141 | C | 12 | 12 | 0.24 pts | 16.24% | 16.31% | earned |
| G142 | C | 12 | 12 | 0.17 pts | 13.03% | 13% | earned |
| G143 | C | 12 | 11 | 0.52 pts | 9.61% | 9.79% | earned |
| G144 | C | 12 | 12 | 0.21 pts | 16.16% | 16.11% | earned |
| G145 | C | 12 | 11 | 0.94 pts | 21.33% | 21.48% | earned |
| G146 | C | 12 | 12 | 0.25 pts | 14.49% | 14.53% | earned |
| G147 | C | 12 | 12 | 0.12 pts | 9.4% | 9.3% | earned |
| G149 | D | 12 | 12 | 0.12 pts | 10.18% | 10.21% | earned |
| G150 | D | 12 | 12 | 0.2 pts | 9.42% | 9.44% | earned |
| G151 | D | 12 | 12 | 0.46 pts | 9.39% | 9.37% | earned |
| G152 | D | 12 | 12 | 0.34 pts | 11.11% | 11.19% | earned |
| G153 | D | 12 | 12 | 0.03 pts | 7.73% | 7.7% | earned |
| G154 | D | 12 | 12 | 0.07 pts | 9.09% | 9.09% | earned |
| G155 | D | 12 | 12 | 0.09 pts | 10.03% | 10.02% | earned |
| G156 | E | 12 | 11 | 0.53 pts | 9.76% | 9.84% | earned |
| G157 | E | 12 | 12 | 0.15 pts | 12.79% | 12.7% | earned |
| G158 | E | 12 | 12 | 0.26 pts | 13.29% | 13.35% | earned |
| G159 | E | 12 | 12 | 0.45 pts | 9.34% | 9.4% | earned |
| G160 | E | 12 | 12 | 0.47 pts | 9.73% | 9.84% | earned |
| G101 | E | 12 | 6 | 0.92 pts | 13.83% | 14.19% | won only 6/12 windows |
| G103 | A | 12 | 6 | 1.64 pts | 12.34% | 12.28% | won only 6/12 windows |
| G110 | A | 12 | 7 | 1.49 pts | 21.79% | 22.08% | won only 7/12 windows |
| G111 | A | 12 | 9 | 0.59 pts | 14.81% | 14.88% | won only 9/12 windows |
| G114 | A | 12 | 6 | 1.01 pts | 13.35% | 13.86% | won only 6/12 windows |
| G121 | A | 12 | 9 | 0.97 pts | 14.97% | 15.1% | won only 9/12 windows |
| G127 | A | 12 | 8 | 0.91 pts | 19.02% | 19.34% | won only 8/12 windows |
| G140 | C | 12 | 9 | 0.54 pts | 15.41% | 15.68% | won only 9/12 windows |
| G148 | D | 12 | 9 | 0.83 pts | 14.2% | 14.45% | won only 9/12 windows |
Event-effect library (median effect on the day, logged events)
| Event | Product | Events | Volume, day of | Volume, day after | Handle time, day of |
|---|---|---|---|---|---|
| weather | A | 16 | +18.42% | +3.45% | +10.54% |
| weather | C | 16 | +20.59% | +3.55% | +10.35% |
| weather | D | 16 | +12.10% | +0.47% | +6.87% |
| weather | E | 16 | +16.07% | 0.00% | +9.70% |
| outage | E | 1 | -52.08% | +25.73% | +32.31% |
What the learning phase changed in the cycle
- Learning rate. The everyday rate dropped to 0.01. At 0.04, corrections chased noise and made the estate worse than the record.
- Shift confirmation. A detector alarm must hold for 7 days (mean error ≥ 0.8σ on the same side) before it counts. Across about 130 series, bare alarms were mostly noise.
- Shrinkage. Migration ratios are shrunk toward plan in proportion to their uncertainty, so a noisy channel does not carry its overshoot into the forecast.
- Earned learning. Learned corrections apply only to gates that earned them (walk-forward, lookback 8 windows, margin 0.25 pts) or had a confirmed shift in the last 70 days.
Synthetic data. In production the forward check runs every day as actuals arrive, and the earning table updates with it.