MacroLens / code /experiments /panel.py
itouchz's picture
Upload experiments/ (runner, predictions, results, paper artifacts)
029e02e verified
Raw
History Blame Contribute Delete
21.2 kB
"""Canonical MacroLens baseline panel for the NeurIPS 2026 D&B submission.
Single source of truth for:
- Which methods are in the panel (18 method classes across 7 families)
- Which tasks each method covers (T1..T7)
- HuggingFace model IDs for LLM/TSFM checkpoints (FP8 native MLLMs)
- GPU parallelism hints (tensor-parallel size)
- Seed strategy (primary seed vs headline T1 subset)
- Ablation subset (5 models x 5 settings on T1 h=21 + T4)
Any change to the panel MUST happen here first; all family runners import from
this module. If a method is not in `ALL_METHODS`, the orchestrators will not
run it. If a HuggingFace ID changes, update this file only.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Literal
# ── Task IDs ──────────────────────────────────────────────────────────────
Task = Literal["T1", "T2", "T3", "T4", "T5", "T6", "T7"]
ALL_TASKS: tuple[Task, ...] = ("T1", "T2", "T3", "T4", "T5", "T6", "T7")
TASK_METADATA: dict[Task, dict] = {
"T1": {"name": "TSF", "long": "Contextual Time-Series Forecasting",
"primary_metric": "MSE"},
"T2": {"name": "Val-PT", "long": "Point-in-Time Equity Valuation",
"primary_metric": "MedAPE"},
"T3": {"name": "Stmt-Gen", "long": "Statement Generation",
"primary_metric": "per-field MAPE"},
"T4": {"name": "Scen-Ret", "long": "Scenario-Conditioned Return Forecasting",
"primary_metric": "Return MAE"},
"T5": {"name": "Priv-Val", "long": "Private-Company Valuation",
"primary_metric": "MedAPE"},
"T6": {"name": "Gen-Eval", "long": "Generator Evaluation",
"primary_metric": "per-field MAPE"},
"T7": {"name": "RE-Val", "long": "Real-Estate Valuation",
"primary_metric": "Rent + Price MAPE"},
}
# ── Method definitions ────────────────────────────────────────────────────
Family = Literal[
"naive", "classical", "sequence",
"tsfm",
"llm_ts",
"llm",
]
@dataclass(frozen=True)
class Method:
"""Single entry in the baseline panel."""
id: str # e.g. "persistence", "chronos2_zs"
name: str # display name, e.g. "Persistence"
family: Family
tasks: frozenset[Task] # tasks this method runs on
hf_id: str | None = None # HuggingFace repo id (for LLM/TSFM)
notes: str = "" # free-form context (size, quant, TP)
# ── Family 1: Naive (4 methods) ───────────────────────────────────────────
# Deterministic heuristics and non-parametric lookups (no fitted parameters).
NAIVE_METHODS: tuple[Method, ...] = (
Method("persistence", "Persistence", "naive",
frozenset({"T1"}),
notes="Repeat last close (T1); repeat pre-event level (T4)."),
Method("sector_median", "Sector-Median", "naive",
frozenset({"T3", "T6"}),
notes="Predict each XBRL field as its sector median."),
Method("metro_median", "Metro-Median", "naive",
frozenset({"T7"}),
notes="Median rent/price in the same metro."),
Method("historical_analogue", "Historical Analogue", "naive",
frozenset({"T4"}),
notes="Find nearest past scenario by type; reuse its post-event return."),
)
# ── Family 2: Classical ML (2 methods) ────────────────────────────────────
# Fitted parametric models (OLS regression, gradient-boosted trees).
CLASSICAL_METHODS: tuple[Method, ...] = (
Method("random_forest", "RandomForest", "classical",
frozenset(ALL_TASKS),
notes="200 trees, max_depth=16, min_samples_leaf=5; sklearn RandomForestRegressor with per-task adapters mirroring LightGBM (log-return target on T1, log-target pipeline on T2/T5/T7, sparse field one-hot on T3/T6, flatten+event-type one-hot on T4)."),
Method("lightgbm", "LightGBM", "classical",
frozenset(ALL_TASKS),
notes=(
"300 trees, num_leaves=63, histogram binning; trained on "
"137-feature panel. Chosen over XGBoost for 2-5x training "
"speedup with essentially identical accuracy on financial "
"tabular data."
)),
)
# ── Family 3: Deep Sequence (3 methods) ───────────────────────────────────
SEQUENCE_METHODS: tuple[Method, ...] = (
Method("dlinear", "DLinear", "sequence",
frozenset({"T1", "T4"}),
notes="Linear decomposition baseline."),
Method("itransformer", "iTransformer", "sequence",
frozenset({"T1", "T4"}),
notes=(
"Inverted transformer (variables-as-tokens); d=128, 4 heads. "
"Chosen over PatchTST as the transformer representative: "
"its cross-variable attention matches the 137-feature "
"multivariate structure of MacroLens better than PatchTST's "
"channel-independent formulation."
)),
Method("moderntcn", "ModernTCN", "sequence",
frozenset({"T1", "T4"}),
notes="Modern pure-convolution backbone."),
)
# ── Family 4: TSFM Zero-Shot (3 methods) ──────────────────────────────────
# Sundial was dropped from the panel because its modeling code (HF Hub
# `thuml/sundial-base-128m`, vendored via `trust_remote_code`) requires
# transformers==4.40.x and is incompatible with transformers>=4.45 (used here
# for vLLM 0.20 + Llama-4 / Gemma-4 / Qwen-3.5 FP8 LLMs); the cascade includes
# DynamicCache.get_usable_length removal, _prepare_4d_causal_attention_mask
# shape mismatch under Sundial's patching, apply_rotary_pos_emb position-id
# scale mismatch, and TSGenerationMixin._extract_past_from_model_output
# removal in GenerationMixin >=4.45. Documented and removed rather than
# patched into a parallel transformers env.
TSFM_ZS_METHODS: tuple[Method, ...] = (
Method("chronos2", "Chronos-2", "tsfm",
frozenset({"T1"}),
hf_id="amazon/chronos-2",
notes="Probabilistic multivariate; frozen checkpoint."),
Method("moirai2", "Moirai 2.0", "tsfm",
frozenset({"T1"}),
hf_id="Salesforce/moirai-2.0-R-small",
notes="Any-variate universal forecaster."),
Method("timesfm", "TimesFM", "tsfm",
frozenset({"T1"}),
hf_id="google/timesfm-1.0-200m-pytorch",
notes=(
"Decoder-only foundation; TimesFM 1.0 (200M, 20 transformer "
"layers). The 2.0 checkpoint (500M, 50 layers) requires a "
"newer `timesfm` package version than the one currently "
"installed; revisit once upgraded."
)),
)
# ── Family 5: LLM-TS Multi-Task (2 methods) ───────────────────────────────
# Note: "LLM-TS Forecasting" family (CALF, TimeReasoner) was removed from the
# panel; the LLM-TS Multi-Task family covers the "LLM adapted for time-series"
# story across all 7 tasks, subsuming the forecast-only variants.
LLM_TS_MULTITASK_METHODS: tuple[Method, ...] = (
Method("chattime", "ChatTime", "llm_ts",
frozenset(ALL_TASKS),
notes="LLaMA-2-7B + 10K-bin tokenisation."),
Method("time_mqa", "Time-MQA", "llm_ts",
frozenset(ALL_TASKS),
notes="Mistral-7B + LoRA r=16; 192,843 QA pairs."),
)
# ── LLM models ────────────────────────────────────────────────────────────
# Paper-canonical Family-6 LLM panel (matches DRAFT.md §5.4 and the
# canon RunRecord JSONs under experiments/results/). Two of the four
# entries are OpenRouter-hosted closed-source models; the third
# (gpt-oss-120B) is open-weights routed via OpenRouter for compute
# economy; the fourth (Qwen-3.5-27B-FP8) runs locally on 4xA100-40GB.
# Local-vLLM fields (tensor_parallel_size, quant, prequantized) are
# meaningful only when ``provider == "local"``; for OpenRouter entries
# they carry placeholder values.
@dataclass(frozen=True)
class LLMModel:
id: str
name: str
provider: Literal["local", "openrouter"]
hf_id: str | None = None
tensor_parallel_size: int = 1
quant: Literal["fp8", "bf16"] = "fp8"
multimodal: bool = False
total_params_b: float | None = None
active_params_b: float | None = None
ft_strategy: str = "none"
prequantized: bool = False
LLM_MODELS: tuple[LLMModel, ...] = (
LLMModel(
id="gpt51",
name="GPT-5.1",
provider="openrouter",
hf_id="openai/gpt-5.1",
ft_strategy="none",
),
LLMModel(
id="gemini3_flash",
name="Gemini-3-Flash-Preview",
provider="openrouter",
hf_id="google/gemini-3-flash-preview",
ft_strategy="none",
),
LLMModel(
id="exaone",
name="EXAONE-4.5 32B",
provider="local",
hf_id="LGAI-EXAONE/EXAONE-4.5-32B-FP8",
tensor_parallel_size=4,
quant="fp8",
total_params_b=32.0,
active_params_b=32.0,
ft_strategy="qlora_nf4",
prequantized=True,
),
LLMModel(
id="llama_scout",
name="Llama-4 Scout 109B",
provider="local",
hf_id="meta-llama/Llama-4-Scout-17B-16E-Instruct",
tensor_parallel_size=4,
quant="fp8",
multimodal=True,
total_params_b=109.0,
active_params_b=17.0,
ft_strategy="qlora_nf4_zero2",
prequantized=False,
),
LLMModel(
id="qwen35",
name="Qwen-3.5-27B-FP8",
provider="local",
hf_id="Qwen/Qwen3.5-27B-FP8",
tensor_parallel_size=1,
quant="fp8",
multimodal=False,
total_params_b=27.0,
active_params_b=27.0,
ft_strategy="qlora_nf4",
prequantized=True,
),
)
LLM_MODELS_BY_ID: dict[str, LLMModel] = {m.id: m for m in LLM_MODELS}
# ── Family 6: LLM Zero-Shot (4 methods) ───────────────────────────────────
# Method ids match the canon RunRecord JSON ``method_id`` field (no
# ``_zs`` suffix); family is ``llm`` (not ``llm_zs``).
def _llm_notes(m: LLMModel) -> str:
if m.provider == "openrouter":
return "OpenRouter API; reasoning tokens disabled."
return f"vLLM {m.quant.upper()} inference, TP={m.tensor_parallel_size}."
LLM_ZS_METHODS: tuple[Method, ...] = tuple(
Method(
id=m.id,
name=m.name,
family="llm",
tasks=frozenset(ALL_TASKS),
hf_id=m.hf_id,
notes=_llm_notes(m),
)
for m in LLM_MODELS
)
# ── Aggregation ───────────────────────────────────────────────────────────
# Paper-canonical 18 methods x 6 families. The legacy ``Method``
# dataclass list aligns with the canon RunRecord JSON ``method_id`` and
# ``method_family`` fields under ``experiments/results/``.
ALL_METHODS: tuple[Method, ...] = (
NAIVE_METHODS
+ CLASSICAL_METHODS
+ SEQUENCE_METHODS
+ TSFM_ZS_METHODS
+ LLM_TS_MULTITASK_METHODS
+ LLM_ZS_METHODS
)
METHODS_BY_ID: dict[str, Method] = {m.id: m for m in ALL_METHODS}
METHODS_BY_FAMILY: dict[Family, tuple[Method, ...]] = {
"naive": NAIVE_METHODS,
"classical": CLASSICAL_METHODS,
"sequence": SEQUENCE_METHODS,
"tsfm": TSFM_ZS_METHODS,
"llm_ts": LLM_TS_MULTITASK_METHODS,
"llm": LLM_ZS_METHODS,
}
def methods_for_task_panel(task: Task) -> tuple[Method, ...]:
"""All legacy panel ``Method`` dataclasses applicable to a task.
Retained under a renamed handle so the new registry-driven
:func:`methods_for_task` (returning ``list[str]`` of registry ids) is
the canonical Phase-4 entry point. Callers that need the panel
dataclass (display name, ``hf_id``, ``notes``) keep using this.
"""
return tuple(m for m in ALL_METHODS if task in m.tasks)
def methods_for_family(family: Family) -> tuple[Method, ...]:
return METHODS_BY_FAMILY[family]
# ── Phase-4 unified-API panel helpers ─────────────────────────────────────
# The orchestrator (``experiments/run_all.py``) consumes the registry-driven
# 18-method panel rather than the legacy ``Method`` dataclasses above. The
# helpers below mirror the registry surface so the runner never reaches into
# ``methods._registry`` directly.
from ..methods._registry import ALL_METHODS as _REGISTRY_METHODS
def methods_for_task(task: str) -> list[str]:
"""Return the sorted list of registered method ids that support ``task``.
Single source of truth for the Phase-4 runner's "skip methods that do
not support this task" filter. Reads directly from
:data:`methods._registry.ALL_METHODS`.
"""
return sorted(name for name, cls in _REGISTRY_METHODS.items() if task in cls.tasks)
# Canonical 18-method panel (re-derived from the registry every call so
# additions/removals show up without an explicit panel.py edit).
PANEL: list[str] = sorted(_REGISTRY_METHODS.keys())
# ── Context ablation subset ──────────────────────────────────────────────
# DRAFT.md §5.4.1: a five-step feature-context ablation (A-E) is run on
# the panel's two zero-shot frontier LLMs (GPT-5.1, Gemini-3-Flash) on
# four tasks (T1 at h=252, T2, T4, T5). Running the full A-E factorial
# across all four LLMs would dominate the wall-clock budget; restricting
# to the two frontier LLMs preserves the contrast (does adding context
# channels help the strongest zero-shot models?) while keeping the
# 2 x 5 x 4 = 40-cell budget tractable.
ABLATION_MODEL_IDS: tuple[str, ...] = ("gpt51", "gemini3_flash")
# The submitted ablation reports zero-shot evaluation only. The FT mode is
# retained as a deferred-experiment slot; with no FT cells the table
# generator (gen_tables.gen_tab_ablation) emits a placeholder.
ABLATION_MODES: tuple[str, ...] = ("ZS",)
# Deferred fine-tune cell (DRAFT.md does not report any FT row in the
# Family-6 panel; the submitted paper is zero-shot-only across all
# four LLMs). Kept as an empty tuple so downstream table generators
# emit the deferred-placeholder branch without crashing.
LLM_FT_PANEL_HF_IDS: tuple[str, ...] = ()
ABLATION_SETTINGS: dict[str, dict] = {
"A": {"name": "OHLCV only", "n_features": 6},
"B": {"name": "A + Fundamentals (XBRL + derived)", "n_features": 70},
"C": {"name": "B + Macro (FRED + EIA)", "n_features": 123},
"D": {"name": "C + Scenario flags", "n_features": 127},
"E": {"name": "D + SBERT filing embeddings", "n_features": 511},
}
ABLATION_TASKS: tuple[Task, ...] = ("T1", "T2", "T4", "T5")
ABLATION_T1_HORIZON: int = 252
# DRAFT.md §5.4.1 / Fig. 3 caption: T1 ablation uses h=252 (the longest
# horizon, where context-channel sensitivity is highest). The other
# three ablation tasks use their full task-defined targets. T3/T6 are
# excluded because they use per-field MAPE / success_rate (different
# metric family); T7 is excluded because RentCast property features
# don't share the A-E feature space (no XBRL / FRED / scenarios).
# ── Seed policy ───────────────────────────────────────────────────────────
# v1 (initial submission): SINGLE seed = 42 for every method, every task.
# Bootstrap 95% CI (1000 resamples) on the test set provides per-method
# variance reporting -- the same approach used by 4 of 7 verified peer
# benchmarks (Time-MMD NeurIPS D&B 2024, FinTSB 2025, Fin-RATE 2026,
# SciTS ICLR 2026), all of which were accepted with single-run headline
# tables. Bootstrap CI captures test-set variance; it does NOT capture
# training-stochasticity variance.
#
# v2 (rebuttal-ready, only fired if reviewer asks): MULTI_SEEDS {42, 123,
# 456} on HEADLINE_T1_MULTISEED_METHODS at T1 h=21. Rationale for matching
# the WIT (ICLR 2026) and EDINET-Bench (ICLR 2026) precedent of 3-run mean
# +/- std on stochastic methods. Estimated rebuttal compute: ~24h on
# 4xA100-40GB (well within the 2-week NeurIPS rebuttal window). Deferring
# to rebuttal saves ~410 GPU-h up front and lets us focus initial
# wall-clock on getting the 20-method panel + 2 deferred FT cells +
# 40-cell ablation factorial fully working at single seed first.
from .. import config as _config
# Single source of truth: the seed lives in config.BENCHMARK_SEED.
# panel.PRIMARY_SEED is kept as the import handle that downstream baselines
# already use, but it MUST stay aligned with config.BENCHMARK_SEED -- the
# assertion below catches any silent drift.
PRIMARY_SEED: int = _config.BENCHMARK_SEED
assert PRIMARY_SEED == _config.BENCHMARK_SEED, (
f"panel.PRIMARY_SEED ({PRIMARY_SEED}) drifted from "
f"config.BENCHMARK_SEED ({_config.BENCHMARK_SEED})"
)
MULTI_SEEDS: tuple[int, ...] = (42, 123, 456)
# Methods that WILL report mean +/- std across MULTI_SEEDS on the headline
# T1 table IF reviewers request multi-seed during rebuttal. List is locked
# in code so the rebuttal path is documented; in v1 the seeds_for() helper
# returns only PRIMARY_SEED.
# Chosen as the stochastic methods present in the current panel; deterministic
# methods (naive, classical without re-sampling, TSFM zero-shot with fixed
# weights) would report a single seed even if multi-seed were enabled.
HEADLINE_T1_MULTISEED_METHODS: tuple[str, ...] = (
"dlinear", "itransformer", "moderntcn",
"time_mqa",
)
# Toggle. v1 = False (single seed everywhere); flip to True during rebuttal
# to activate multi-seed for HEADLINE_T1_MULTISEED_METHODS at T1 h=21.
ENABLE_MULTI_SEED: bool = False
def seeds_for(method_id: str, task: Task, horizon: int | None = None) -> tuple[int, ...]:
"""Return the seed list for a method-task pair.
v1 (initial submission, ENABLE_MULTI_SEED=False): always returns
(PRIMARY_SEED,) -- single seed everywhere.
v2 (rebuttal, ENABLE_MULTI_SEED=True): returns MULTI_SEEDS on the
headline T1 subset (method in HEADLINE_T1_MULTISEED_METHODS,
task == 'T1', horizon == 21); single seed otherwise.
"""
if (
ENABLE_MULTI_SEED
and task == "T1"
and horizon == 21
and method_id in HEADLINE_T1_MULTISEED_METHODS
):
return MULTI_SEEDS
return (PRIMARY_SEED,)
# ── GPU assignment ────────────────────────────────────────────────────────
# MacroLens runs on GPU IDs 4,5,6,7 on the shared host (last 4 of the 8
# physical A100-SXM4-40GB). All scripts must respect this;
# `CUDA_VISIBLE_DEVICES` is set by the runner wrappers.
# (Memory: project_macrolens_gpus.md)
GPU_IDS: tuple[int, ...] = (4, 5, 6, 7)
CUDA_VISIBLE_DEVICES_STR: str = ",".join(str(i) for i in GPU_IDS)
# ── Summary ───────────────────────────────────────────────────────────────
def summary() -> dict:
"""Return a small dict summarising the panel for logging / CI assertions."""
return {
"total_methods": len(ALL_METHODS),
"per_family": {f: len(ms) for f, ms in METHODS_BY_FAMILY.items()},
"per_task": {t: len(methods_for_task_panel(t)) for t in ALL_TASKS},
"ablation_models": len(ABLATION_MODEL_IDS),
"ablation_settings": len(ABLATION_SETTINGS),
"gpu_ids": list(GPU_IDS),
"primary_seed": PRIMARY_SEED,
"multi_seeds": list(MULTI_SEEDS),
"llm_models": [m.hf_id for m in LLM_MODELS],
}
if __name__ == "__main__":
# Quick sanity check: `python -m baselines.panel`.
# Expected total: 18 methods x 6 families (4 naive + 2 classical
# + 3 sequence + 3 tsfm + 2 llm_ts + 4 llm) -- matches DRAFT.md §5.4
# and the canon RunRecord JSONs under experiments/results/.
import json
s = summary()
assert s["total_methods"] == 18, f"Expected 18 methods, got {s['total_methods']}"
assert len(s["per_family"]) == 6, (
f"Expected 6 families, got {len(s['per_family'])}"
)
print(json.dumps(s, indent=2, default=list))
print(
f"Context ablation: 2 frontier LLMs x A-E x {{T1 h={ABLATION_T1_HORIZON},"
f" T2, T4, T5}} = 2 x {len(ABLATION_SETTINGS)}"
f" x {len(ABLATION_TASKS)} ="
f" {2 * len(ABLATION_SETTINGS) * len(ABLATION_TASKS)} cells"
" (DRAFT.md §5.4.1)."
)