Skip to content
everesteer

Theme
Sign inSign up
FIXED ROUTES

Benchmark models.

Reference implementations the tournament publishes. Climb past them to earn alpha. Open-source code, deterministic outputs, weekly updates.

Scored benchmarks · updated weekly24 models
ModelArchTargetMean CORRMean AIMCAvg payoutSource
lgb_cho_oyuLGBtarget_everest_200.04590.0428+0.0190
lgb_dartLGBtarget_everest_200.06260.0358+0.0175
v0_sherpa_8V0target_everest_20-0.01140.0160+0.0167
v0_sherpa_7V0target_everest_20-0.01030.0133+0.0167
mlp_primaryMLPtarget_everest_200.06080.0237+0.0153
ridge_primaryRIDGEtarget_everest_200.06770.0360+0.0136
elasticnet_primaryELASTICNETtarget_everest_200.11660.1057+0.0125
xgb_3targetXGBtarget_everest_200.02610.0168+0.0052
lgb_3targetLGBtarget_everest_20-0.0071-0.0154+0.0017
lgb_multitaskLGBtarget_everest_20-0.0007-0.0188-0.0027
rf_primaryRFtarget_everest_200.0194-0.0375-0.0125
lgb_kangchenLGBtarget_everest_200.0190-0.0144-0.0125
xgb_primaryXGBtarget_everest_200.0079-0.0301-0.0125
v0_sherpa_3V0target_everest_20-0.0481-0.0245-0.0167
v0_sherpa_9V0target_everest_20-0.0645-0.0511-0.0167
v0_sherpa_5V0target_everest_20-0.0205-0.0068-0.0167
v0_sherpa_10V0target_everest_20-0.0350-0.0342-0.0167
v0_sherpa_2V0target_everest_20-0.0203-0.0018-0.0167
v0_sherpa_12V0target_everest_20-0.0682-0.0707-0.0167
v0_sherpa_1V0target_everest_20-0.0297-0.0124-0.0167
v0_sherpa_11V0target_everest_20-0.0590-0.0550-0.0167
v0_sherpa_4V0target_everest_20-0.0995-0.1192-0.0167
v0_sherpa_6V0target_everest_20-0.0539-0.0560-0.0167
lgb_primaryLGBtarget_everest_20-0.0892-0.0867-0.0273
Performance · top model vs benchmarks

Round-by-round α

Benchmark time-series will appear as rounds resolve.

Mechanics

Benchmark models produce fully deterministic outputs. Given the same live feature set, every run returns identical predictions. No random seeds, no cloud state, no stale caches. You can reproduce any round's benchmark score from the public repository with a single command.

All benchmark code is published on GitHub under the Everesteer organisation before each season opens. The repository includes training scripts, frozen weight files, and a validation harness that matches the tournament's internal scorer.

"Beating" a benchmark means your model's tournament-settled alpha exceeds the benchmark's alpha in the same round, net of any correlation penalty. The tournament tracks a 20-round rolling beat-rate. Models above 60% against eiq_minera_model qualify for the Top 14 consideration pool.

Benchmarks update weekly on the Wednesday before each round opens. The update pulls the latest live feature set, re-runs inference, and commits new predictions to the repository within two hours. Benchmark scores appear in the leaderboard alongside participant models.

The eiq_minera_model benchmark uses stake as weight: each of the top 14 models' predictions contributes proportionally to its current stake balance. This makes it a live measure of collective conviction rather than a static average.

To propose a new benchmark for inclusion, open a pull request to everestquant/benchmarks with a deterministic inference script, a frozen weight artifact under 50 MB, and a backtest report covering at least 100 rounds.

Think you can beat them?