TW3Cast
A time-series forecasting system based on a statically frozen routing table. Routing strategies determined during training select among multiple forecasting modes. The system combines LoRA fine-tuning with ensemble methods to achieve high-accuracy time-series forecasting without requiring agents or large language models.
Timeline (last 90 days)
TW3Cast ranks 3rd on the GIFT-Eval time series forecasting benchmark (out of 130 entries, ranked by average MASE), with the ranking snapshot dated 2026-09-14.
TW3Cast uses no agents and does not call any language models.
The two GIFT-Eval entries ranked above TW3Cast belong to the agentic category of the leaderboard.
The core of TW3Cast is a routing table computed once on the training set and then completely frozen, with no dynamic runtime decision-making.
TW3Cast faces 97 combinations of dataset, frequency, and forecast horizon configurations on GIFT-Eval, and the routing table selects one of four modes for each configuration.
TW3Cast applies LoRA or full fine-tuning to three publicly available time series foundation models: Chronos-2, TiRex, and Toto.
Every decision in TW3Cast's routing table is made on a backtest set held out from the training set, without touching the test set at all.
TW3Cast's candidate admission mechanism: an expert model is immediately accepted as long as it outperforms the current result on the backtest tournament, and trying a new candidate requires only a few megabytes of storage and a few minutes of GPU time.
TW3Cast uses a three-guard mechanism to prevent selection bias: dual criteria of accuracy and calibration, targeted asymmetric margins, and conservative per-window gating.
TW3Cast's selection rules themselves are chosen through a temporal meta-backtest, forming nested validation.
3 more timeline events
All Facts (11)
TW3Cast uses no agents and does not call any language models.
50%UnverifiedThe core of TW3Cast is a routing table computed once on the training set and then completely frozen, with no dynamic runtime decision-making.
50%UnverifiedTW3Cast faces 97 combinations of dataset, frequency, and forecast horizon configurations on GIFT-Eval, and the routing table selects one of four modes for each configuration.
50%UnverifiedTW3Cast applies LoRA or full fine-tuning to three publicly available time series foundation models: Chronos-2, TiRex, and Toto.
50%UnverifiedEvery decision in TW3Cast's routing table is made on a backtest set held out from the training set, without touching the test set at all.
50%UnverifiedTW3Cast's candidate admission mechanism: an expert model is immediately accepted as long as it outperforms the current result on the backtest tournament, and trying a new candidate requires only a few megabytes of storage and a few minutes of GPU time.
50%UnverifiedTW3Cast uses a three-guard mechanism to prevent selection bias: dual criteria of accuracy and calibration, targeted asymmetric margins, and conservative per-window gating.
50%UnverifiedTW3Cast's selection rules themselves are chosen through a temporal meta-backtest, forming nested validation.
50%UnverifiedComparison data from the paper: the best individual foundation model alone has an average MASE rank of 33.8, per-configuration tournament selection is 38.0, and the full router is 19.4.
50%UnverifiedPure tournament selection (38.0) performs worse than the best single model (33.8), indicating that per-configuration selection suffers from overfitting the backtest.
50%UnverifiedTW3Cast makes public its routing table, expert index, fixed-version base models, submitted score files, and dated snapshots of public scores, with every leaderboard number in the paper reproducible via a single script.
50%