Symmetric-averaging numbers (paper Table 2); Timestral restricted licence; fev-bench pairwise

README.md

@@ -11,6 +11,8 @@ configs:
     default: true
   - config_name: tasks
     data_files: tasks.csv
+  - config_name: pairwise
+    data_files: pairwise.csv
 ---
 
 # RealCast-bench Leaderboard
@@ -18,7 +20,7 @@ configs:
 Time series foundation models evaluated zero-shot on proprietary Carbon Arc
 commercial panels: 10 source panels, 81 datasets, 290 tasks across three boards
 (152 weekly, 102 monthly, 36 daily). Test year: October 2025 to September 2026.
-This table mirrors Table 4 of the paper *RealCast-bench: Evaluating Time Series Foundation Models on Proprietary Commercial Panels*. The panels themselves are
+This table mirrors Table 2 of the paper *RealCast-bench: Evaluating Time Series Foundation Models on Proprietary Commercial Panels*. The panels themselves are
 proprietary and are not released.
 
 ## Columns
@@ -32,11 +34,10 @@ proprietary and are not released.
 - **`*_rank`**: rank among the 21 configurations.
 - **`win_rate`**: share of (task, other configuration) pairings in which this
   configuration has the lower MASE, with ties counted as half.
-- **`tier`**: configurations are taken in order of overall score. A new tier starts
-  when a configuration is significantly worse than the first configuration of the
-  current tier (paired entity bootstrap, 1,000 resamples, Holm correction). The
-  tiers are finer than a rerun would support, so treat a 1% difference as the
-  practical cut.
+- **`license`**, **`license_url`**: the license of the released weights, with a link to
+  its text.
+- **`commercial_use`**: `yes` (Apache-2.0), `no` (non-commercial license) or
+  `restricted` (commercial use permitted under conditions; see the license).
 - **`covariates`**: `none` means the model saw only the target series. `past +
   calendar` means it also saw the panel's other measured series up to the forecast
   origin, plus calendar features.
@@ -56,9 +57,25 @@ AutoETS's WQL explodes on 11 tasks (up to about 1e35, worst on daily
 prediction intervals. These values are reported as computed. The leaderboard
 ranks by MASE, which they do not affect.
 
+## Pairwise comparison (`pairwise` config)
+
+Pairwise win rate and skill score between configurations, computed from `tasks.csv` with
+[fev-bench](https://github.com/autogluon/fev)'s method (`fev.analysis.pairwise_comparison`,
+commit f9e2c7d, reimplemented; it reproduces fev's output exactly). One row per (board, metric, model_1, model_2); `board` is `all`, `weekly`,
+`monthly` or `daily`, and `metric` is `mase` or `wql`. `win_rate` is the share of tasks on
+which `model_1` has the lower error than `model_2`, ties counted as half. `skill_score` is 1
+minus the geometric mean over tasks of `model_1`'s error divided by `model_2`'s, with ratios
+clipped to [0.01, 100]. `_lower` and `_upper` bound the 95% interval from 1,000 bootstrap
+resamples of tasks (seed 123). As on fev-bench, each (board, metric) holds the 15
+configurations with the highest average win rate, plus Seasonal Naive. `+ cov.` means past +
+calendar covariates.
+
 ## Setup
 
-- No model is fine-tuned on Carbon Arc data, and all run with default settings.
+- No model is fine-tuned on Carbon Arc data.
+- Symmetric averaging (averaging the forecast of a series with the sign-restored forecast
+  of its negation) is enabled for every model that offers it: TimesFM-2.5, TimesFM-3.0,
+  Timestral-209M and the three t0 models. Other settings are each library's defaults.
 
 ## Submitting a model
 

leaderboard.csv

@@ -1,22 +1,22 @@
-overall_rank,configuration,overall_rel_mase,win_rate,tier,weekly_rel_mase,weekly_rank,monthly_rel_mase,monthly_rank,daily_rel_mase,daily_rank,model,covariates,organization,params,model_type,license,commercial_use,note
-1,Timestral-209M + past & calendar covariates,0.553,0.89,1,0.521,1,0.564,1,0.673,1,Timestral-209M,past + calendar,Mistral AI,209M,Pretrained,Modified MIT (no rights above $20M monthly company revenue),False,
-2,Timestral-209M,0.557,0.838,2,0.526,3,0.567,2,0.678,2,Timestral-209M,none,Mistral AI,209M,Pretrained,Modified MIT (no rights above $20M monthly company revenue),False,
-3,TimesFM-3.0 + past & calendar covariates,0.558,0.801,2,0.524,2,0.569,3,0.688,5,TimesFM-3.0,past + calendar,Google,300M,Pretrained,TimesFM Non-Commercial v1.0,False,
-4,TimesFM-3.0,0.563,0.724,3,0.526,4,0.578,5,0.692,7,TimesFM-3.0,none,Google,300M,Pretrained,TimesFM Non-Commercial v1.0,False,
-5,t0-gamma + past & calendar covariates,0.563,0.736,3,0.528,5,0.576,4,0.69,6,t0-gamma,past + calendar,The Forecasting Company,256M,Pretrained,Apache-2.0,True,
-6,t0-gamma,0.571,0.595,4,0.538,8,0.585,8,0.688,3,t0-gamma,none,The Forecasting Company,256M,Pretrained,Apache-2.0,True,
-7,Chronos-2 + past & calendar covariates,0.572,0.568,4,0.537,7,0.584,7,0.7,8,Chronos-2,past + calendar,AWS,120M,Pretrained,Apache-2.0,True,
-8,PatchTST-FM-r2,0.572,0.597,4,0.531,6,0.599,15,0.688,4,PatchTST-FM-r2,none,IBM,385M,Pretrained,Apache-2.0,True,
-9,Chronos-2,0.575,0.521,5,0.54,11,0.589,12,0.7,10,Chronos-2,none,AWS,120M,Pretrained,Apache-2.0,True,
-10,TimesFM-2.5,0.576,0.541,5,0.539,9,0.593,14,0.701,11,TimesFM-2.5,none,Google,200M,Pretrained,Apache-2.0,True,
-11,t0-beta + past & calendar covariates,0.576,0.554,5,0.54,10,0.589,11,0.716,15,t0-beta,past + calendar,The Forecasting Company,256M,Pretrained,Apache-2.0,True,
-12,Toto-2.0-22M + past & calendar covariates,0.578,0.522,5,0.55,13,0.578,6,0.707,12,Toto-2.0-22M,past + calendar,Datadog,22M,Pretrained,Apache-2.0,True,experimental covariates (vendor: not yet supported)
-13,Toto-2.0-22M,0.582,0.458,6,0.551,14,0.588,10,0.708,13,Toto-2.0-22M,none,Datadog,22M,Pretrained,Apache-2.0,True,
-14,TiRex-2 + past & calendar covariates,0.584,0.49,6,0.554,15,0.587,9,0.717,16,TiRex-2,past + calendar,NX-AI,83M,Pretrained,Apache-2.0,True,
-15,t0-beta,0.585,0.407,6,0.55,12,0.601,16,0.7,9,t0-beta,none,The Forecasting Company,256M,Pretrained,Apache-2.0,True,
-16,TiRex-2,0.586,0.439,6,0.555,16,0.59,13,0.718,17,TiRex-2,none,NX-AI,83M,Pretrained,Apache-2.0,True,
-17,t0-alpha,0.599,0.312,7,0.564,18,0.615,18,0.711,14,t0-alpha,none,The Forecasting Company,102M,Pretrained,Apache-2.0,True,
-18,t0-alpha + past & calendar covariates,0.6,0.276,7,0.564,17,0.613,17,0.732,18,t0-alpha,past + calendar,The Forecasting Company,102M,Pretrained,Apache-2.0,True,
-19,AutoTheta,0.678,0.133,8,0.637,19,0.68,20,0.871,20,AutoTheta,none,Nixtla (StatsForecast),,Statistical,Apache-2.0,True,
-20,AutoETS,0.692,0.091,9,0.667,20,0.677,19,0.862,19,AutoETS,none,Nixtla (StatsForecast),,Statistical,Apache-2.0,True,
-21,Seasonal Naive,1.0,0.007,10,1.0,21,1.0,21,1.0,21,Seasonal Naive,none,Nixtla (StatsForecast),,Statistical,Apache-2.0,True,
+overall_rank,configuration,overall_rel_mase,win_rate,weekly_rel_mase,weekly_rank,monthly_rel_mase,monthly_rank,daily_rel_mase,daily_rank,model,covariates,organization,params,model_type,license,commercial_use,license_url,note
+1,Timestral-209M + past & calendar covariates,0.553,0.881,0.521,1,0.564,1,0.673,1,Timestral-209M,past + calendar,Mistral AI,209M,Pretrained,Modified MIT,restricted,https://huggingface.co/mistralai/Timestral-209M,
+2,TimesFM-3.0 + past & calendar covariates,0.556,0.828,0.523,2,0.566,2,0.684,3,TimesFM-3.0,past + calendar,Google,300M,Pretrained,TimesFM Non-Commercial v1.0,no,https://huggingface.co/google/timesfm-3.0-pytorch,
+3,Timestral-209M,0.557,0.825,0.526,4,0.567,3,0.678,2,Timestral-209M,none,Mistral AI,209M,Pretrained,Modified MIT,restricted,https://huggingface.co/mistralai/Timestral-209M,
+4,TimesFM-3.0,0.561,0.75,0.525,3,0.575,5,0.688,6,TimesFM-3.0,none,Google,300M,Pretrained,TimesFM Non-Commercial v1.0,no,https://huggingface.co/google/timesfm-3.0-pytorch,
+5,t0-gamma + past & calendar covariates,0.562,0.738,0.527,5,0.574,4,0.691,7,t0-gamma,past + calendar,The Forecasting Company,256M,Pretrained,Apache-2.0,yes,,
+6,t0-gamma,0.568,0.612,0.536,7,0.58,7,0.687,4,t0-gamma,none,The Forecasting Company,256M,Pretrained,Apache-2.0,yes,,
+7,Chronos-2 + past & calendar covariates,0.572,0.551,0.537,9,0.584,8,0.7,8,Chronos-2,past + calendar,AWS,120M,Pretrained,Apache-2.0,yes,,
+8,PatchTST-FM-r2,0.572,0.58,0.531,6,0.599,16,0.688,5,PatchTST-FM-r2,none,IBM,385M,Pretrained,Apache-2.0,yes,,
+9,t0-beta + past & calendar covariates,0.574,0.574,0.536,8,0.587,9,0.715,15,t0-beta,past + calendar,The Forecasting Company,256M,Pretrained,Apache-2.0,yes,,
+10,Chronos-2,0.575,0.501,0.54,11,0.589,12,0.7,10,Chronos-2,none,AWS,120M,Pretrained,Apache-2.0,yes,,
+11,TimesFM-2.5,0.576,0.523,0.539,10,0.593,14,0.701,11,TimesFM-2.5,none,Google,200M,Pretrained,Apache-2.0,yes,,
+12,Toto-2.0-22M + past & calendar covariates,0.578,0.503,0.55,13,0.578,6,0.707,12,Toto-2.0-22M,past + calendar,Datadog,22M,Pretrained,Apache-2.0,yes,,experimental covariates (vendor: not yet supported)
+13,t0-beta,0.58,0.445,0.545,12,0.596,15,0.7,9,t0-beta,none,The Forecasting Company,256M,Pretrained,Apache-2.0,yes,,
+14,Toto-2.0-22M,0.582,0.44,0.551,14,0.588,11,0.708,13,Toto-2.0-22M,none,Datadog,22M,Pretrained,Apache-2.0,yes,,
+15,TiRex-2 + past & calendar covariates,0.584,0.466,0.554,15,0.587,10,0.717,16,TiRex-2,past + calendar,NX-AI,83M,Pretrained,Apache-2.0,yes,,
+16,TiRex-2,0.586,0.419,0.555,16,0.59,13,0.718,17,TiRex-2,none,NX-AI,83M,Pretrained,Apache-2.0,yes,,
+17,t0-alpha,0.591,0.354,0.557,17,0.605,17,0.709,14,t0-alpha,none,The Forecasting Company,102M,Pretrained,Apache-2.0,yes,,
+18,t0-alpha + past & calendar covariates,0.597,0.285,0.562,18,0.608,18,0.729,18,t0-alpha,past + calendar,The Forecasting Company,102M,Pretrained,Apache-2.0,yes,,
+19,AutoTheta,0.678,0.128,0.637,19,0.68,20,0.871,20,AutoTheta,none,Nixtla (StatsForecast),,Statistical,Apache-2.0,yes,,
+20,AutoETS,0.692,0.089,0.667,20,0.677,19,0.862,19,AutoETS,none,Nixtla (StatsForecast),,Statistical,Apache-2.0,yes,,
+21,Seasonal Naive,1.0,0.007,1.0,21,1.0,21,1.0,21,Seasonal Naive,none,Nixtla (StatsForecast),,Statistical,Apache-2.0,yes,,