You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

436 lines
16 KiB
Python

"""SQS computation and scoring API endpoints."""
from __future__ import annotations
import threading
import uuid
from pathlib import Path
from typing import Any
from fastapi import APIRouter, HTTPException
from pydantic import BaseModel
from apps.web.dependencies import get_configs_dir, get_journal_dir, get_runs_dir
from libs.backtest.tracker import (
_hydrate_split_result,
compute_public_sqs_v9,
compute_rqs,
compute_wfqs_v2,
load_journal,
refresh_public_scores,
)
router = APIRouter(prefix="/sqs", tags=["sqs"])
# ---------------------------------------------------------------------------
# In-memory pipeline state
# ---------------------------------------------------------------------------
_pipelines: dict[str, dict[str, Any]] = {}
_pipelines_lock = threading.Lock()
class RescoreRequest(BaseModel):
target: str
class PipelineRequest(BaseModel):
target: str
_NO_ENTRY_RESPONSE = {
"entry_id": None,
"sqs_score": None,
"source": None,
"pending_reason": "no_backtest_data",
"pillars": {
"rqs_score": None, "wfqs_v2_score": None,
"regime_score": None, "regime_source": "none", "core_score": None,
},
"gates": {
"deployment_gate_factor": None,
"robustness_gate_factor": None,
"activity_factor": None,
},
"breakdown": {},
"stored_sqs_score": None,
"scenario_robustness_score": None,
"overfit_check_score": None,
}
def _registry_fallback(journal_dir: Path, experiment_name: str) -> dict[str, Any] | None:
"""Look up stored SQS from registry when no live journal entry exists."""
from libs.backtest.domain import ExperimentRegistry
registry_path = journal_dir / "experiment_registry.json"
if not registry_path.exists():
return None
try:
registry = ExperimentRegistry.model_validate_json(registry_path.read_text())
except Exception:
return None
matches = [e for e in registry.entries if e.experiment_name == experiment_name]
if not matches:
return None
entry = sorted(matches, key=lambda e: e.timestamp)[-1]
return {
"entry_id": entry.entry_id,
"sqs_score": entry.sqs_score,
"source": "registry_stored",
"pending_reason": None if entry.sqs_score is not None else "no_backtest_data",
"pillars": {
"rqs_score": entry.rqs_score,
"wfqs_v2_score": entry.wfqs_v2_score,
"regime_score": None,
"regime_source": "stored",
"core_score": None,
},
"gates": {
"deployment_gate_factor": None,
"robustness_gate_factor": None,
"activity_factor": None,
},
"breakdown": {},
"stored_sqs_score": entry.sqs_score,
"scenario_robustness_score": entry.scenario_robustness_score,
"overfit_check_score": entry.overfit_check_score,
}
def _compute_for_entry(entry: Any) -> dict[str, Any]:
"""Run SQS v9 computation on a journal entry and return structured result."""
train_result = _hydrate_split_result(entry.results.get("train"))
valid_result = _hydrate_split_result(entry.results.get("valid"))
test_result = _hydrate_split_result(entry.results.get("test"))
rqs_score, _ = compute_rqs(train_result, valid_result, test_result)
wfqs_v2_score, _ = compute_wfqs_v2(entry.walk_forward_summary)
sqs_score, sqs_breakdown, source = compute_public_sqs_v9(
train_result, valid_result, test_result,
walk_forward_summary=entry.walk_forward_summary,
robustness_matrix_summary=entry.robustness_matrix_summary,
out_of_time_robustness_summary=entry.out_of_time_robustness_summary,
common_window_summary=entry.common_window_summary,
reset_common_window_summary=entry.reset_common_window_summary,
scenario_robustness_score=entry.scenario_robustness_score,
rqs_score=rqs_score,
wfqs_v2_score=wfqs_v2_score,
)
regime_src = (
"scenario_rrs" if entry.scenario_robustness_score is not None
else "oot_quality" if entry.out_of_time_robustness_summary is not None
else "neutral_50"
)
pending_reason = None
if sqs_score is None:
pending_key = next((k for k in sqs_breakdown if k.startswith("requires_")), None)
pending_reason = (
pending_key.replace("requires_", "").replace("_", " ")
if pending_key else "missing data"
)
return {
"experiment_name": entry.experiment_name,
"entry_id": entry.entry_id,
"sqs_score": sqs_score,
"source": source,
"pending_reason": pending_reason,
"pillars": {
"rqs_score": rqs_score,
"wfqs_v2_score": wfqs_v2_score,
"regime_score": sqs_breakdown.get("regime_score"),
"regime_source": regime_src,
"core_score": sqs_breakdown.get("core_score"),
},
"gates": {
"deployment_gate_factor": sqs_breakdown.get("deployment_gate_factor"),
"robustness_gate_factor": sqs_breakdown.get("gate_factor"),
"activity_factor": sqs_breakdown.get("activity_factor"),
},
"breakdown": sqs_breakdown,
"stored_sqs_score": entry.sqs_score,
"scenario_robustness_score": entry.scenario_robustness_score,
"overfit_check_score": entry.overfit_check_score,
}
@router.get("/{experiment_name:path}")
def get_sqs(experiment_name: str) -> dict[str, Any]:
"""Compute SQS v9 live for an experiment (read-only, no journal write).
Falls back to registry-stored SQS when no journal entry exists.
Never returns 404 — always returns a structured response.
"""
journal_dir = get_journal_dir()
journal_path = journal_dir / "improvement_journal.jsonl"
base = {"experiment_name": experiment_name}
# 1. Try journal — live recompute
if journal_path.exists():
entries = load_journal(journal_path)
exact = [e for e in entries if e.experiment_name == experiment_name]
if exact:
entry = sorted(exact, key=lambda e: e.timestamp)[-1]
return {**base, **_compute_for_entry(entry)}
# 2. Fall back to registry (stored values, no live recompute)
reg = _registry_fallback(journal_dir, experiment_name)
if reg:
return {**base, **reg}
# 3. Nothing available
return {**base, **_NO_ENTRY_RESPONSE}
def _run_pipeline_thread(pipeline_id: str, experiment_name: str) -> None:
"""Background thread: run train/valid/test backtests → record → compute SQS."""
import datetime as _dt
from libs.backtest.tracker import (
append_journal_entry,
build_split_result,
get_next_entry_id,
journal_lock,
rebuild_registry,
scan_runs_for_experiment,
)
from libs.backtest.domain import JournalEntry
def _set(updates: dict) -> None:
with _pipelines_lock:
_pipelines[pipeline_id].update(updates)
try:
configs_dir = get_configs_dir()
journal_dir = get_journal_dir()
runs_dir = get_runs_dir()
journal_path = journal_dir / "improvement_journal.jsonl"
registry_path = journal_dir / "experiment_registry.json"
leaderboard_path = journal_dir / "LEADERBOARD.md"
config_path = configs_dir / f"{experiment_name}.json"
if not config_path.exists():
_set({"status": "failed", "error": f"Config not found: {experiment_name}.json"})
return
# Import backtester internals
from apps.backtester.run import (
BacktestRunner,
_build_merged_snapshot_store,
_extend_store_to_requested_window,
load_manifest,
resolve_config,
)
from libs.backtest.snapshots import resolve_snapshot_path
import pyarrow.parquet as _pq
_set({"status": "running", "step": "Loading snapshot store…"})
manifest = load_manifest(str(config_path))
config = resolve_config(manifest)
# Build merged snapshot store — uses cached PKL, completes in seconds
store = _build_merged_snapshot_store(manifest, config, snapshot_dir_override=None)
# Determine split date ranges from parquet files (read only execution_date column)
snapshot_request_id = config.requested_snapshot_id or config.dataset_snapshot_id
snapshot_dir = resolve_snapshot_path(snapshot_request_id)
if snapshot_dir is None:
_set({"status": "failed", "error": f"Snapshot directory not found for '{snapshot_request_id}'"})
return
split_date_ranges: dict[str, tuple[_dt.date, _dt.date]] = {}
for split in ("train", "valid", "test"):
parquet_path = snapshot_dir / f"{split}.parquet"
if not parquet_path.exists():
_set({"status": "failed", "error": f"Parquet not found: {parquet_path}"})
return
tbl = _pq.read_table(str(parquet_path), columns=["execution_date"])
raw_dates = tbl["execution_date"].to_pylist()
dates: list[_dt.date] = []
for v in raw_dates:
if v is None:
continue
if isinstance(v, _dt.datetime):
dates.append(v.date())
elif isinstance(v, _dt.date):
dates.append(v)
elif isinstance(v, str):
dates.append(_dt.date.fromisoformat(v[:10]))
if not dates:
_set({"status": "failed", "error": f"No execution dates in {split}.parquet"})
return
split_date_ranges[split] = (min(dates), max(dates))
# Run each split in-process using the cached merged store (slice per split)
for split in ("train", "valid", "test"):
_set({"status": "running", "step": f"Running {split} split…"})
split_start, split_end = split_date_ranges[split]
split_store = store.slice_by_date_range(split_start, split_end)
split_store = _extend_store_to_requested_window(
store=split_store,
config=config,
start_date=split_start,
end_date=split_end,
snapshot_dir_override=None,
)
runner = BacktestRunner(
manifest=manifest,
config=config,
store=split_store,
initial_equity=10_000.0,
split_name=split,
)
runner.run(output_root=str(runs_dir))
# Record
_set({"status": "running", "step": "Recording results to journal…"})
split_runs = scan_runs_for_experiment(runs_dir, experiment_name)
if not split_runs:
_set({"status": "failed", "error": "Backtests completed but no runs found in runs/. Check output-root."})
return
results: dict[str, Any] = {}
for sname, (run_id, metrics) in split_runs.items():
results[sname] = build_split_result(sname, run_id, metrics)
rqs_score, rqs_breakdown = compute_rqs(results.get("train"), results.get("valid"), results.get("test"))
sqs_score, sqs_breakdown, _ = compute_public_sqs_v9(
results.get("train"), results.get("valid"), results.get("test"),
rqs_score=rqs_score, wfqs_v2_score=None,
)
tags = [t for t in experiment_name.replace("-", "_").split("_") if t]
entry_id = get_next_entry_id(journal_path)
timestamp = _dt.datetime.now(_dt.timezone.utc).isoformat()
entry = JournalEntry(
entry_id=entry_id,
timestamp=timestamp,
experiment_name=experiment_name,
hypothesis="Auto-recorded via web GUI",
results=results,
rqs_score=rqs_score,
rqs_breakdown=rqs_breakdown,
sqs_score=sqs_score,
sqs_breakdown=sqs_breakdown,
verdict="unknown",
tags=tags,
)
with journal_lock(journal_path):
append_journal_entry(journal_path, entry)
rebuild_registry(journal_path, registry_path, leaderboard_path)
# Compute final SQS
_set({"status": "running", "step": "Computing SQS v9…"})
fresh_entries = load_journal(journal_path)
fresh_exact = [e for e in fresh_entries if e.experiment_name == experiment_name]
fresh_entry = sorted(fresh_exact, key=lambda e: e.timestamp)[-1]
sqs_result = _compute_for_entry(fresh_entry)
_set({"status": "completed", "step": "Done", "result": sqs_result})
except Exception as exc:
_set({"status": "failed", "error": str(exc)})
@router.post("/pipeline")
def pipeline_start(req: PipelineRequest) -> dict[str, Any]:
"""Start the auto-backtest pipeline: run train/valid/test → record → compute SQS.
Returns a pipeline_id to poll with GET /sqs/pipeline/{pipeline_id}.
If a journal entry already exists, computes SQS immediately (no pipeline).
"""
experiment_name = req.target
journal_dir = get_journal_dir()
journal_path = journal_dir / "improvement_journal.jsonl"
base = {"experiment_name": experiment_name}
# Journal entry already exists → compute immediately
if journal_path.exists():
entries = load_journal(journal_path)
exact = [e for e in entries if e.experiment_name == experiment_name]
if exact:
entry = sorted(exact, key=lambda e: e.timestamp)[-1]
return {**base, "pipeline_status": "existing_entry", **_compute_for_entry(entry)}
# Config must exist before we start
config_path = get_configs_dir() / f"{experiment_name}.json"
if not config_path.exists():
raise HTTPException(status_code=404, detail=f"Experiment config not found: {experiment_name}")
pipeline_id = str(uuid.uuid4())[:8]
with _pipelines_lock:
_pipelines[pipeline_id] = {
"pipeline_id": pipeline_id,
"experiment_name": experiment_name,
"status": "running",
"step": "Starting…",
"error": None,
"result": None,
}
thread = threading.Thread(
target=_run_pipeline_thread,
args=(pipeline_id, experiment_name),
daemon=True,
)
thread.start()
return {**base, "pipeline_id": pipeline_id, "pipeline_status": "running", "step": "Starting…"}
@router.get("/pipeline/{pipeline_id}")
def pipeline_poll(pipeline_id: str) -> dict[str, Any]:
"""Poll pipeline status. Returns result when status='completed'."""
with _pipelines_lock:
state = _pipelines.get(pipeline_id)
if state is None:
raise HTTPException(status_code=404, detail=f"Pipeline not found: {pipeline_id}")
out = dict(state)
if out.get("result"):
out.update(out.pop("result"))
out["pipeline_status"] = out.pop("status")
return out
@router.post("/rescore")
def rescore_experiment(req: RescoreRequest) -> dict[str, Any]:
"""Recompute SQS for one experiment, save to journal, rebuild leaderboard."""
from libs.backtest.tracker import journal_lock, rebuild_registry
journal_dir = get_journal_dir()
journal_path = journal_dir / "improvement_journal.jsonl"
registry_path = journal_dir / "experiment_registry.json"
leaderboard_path = journal_dir / "LEADERBOARD.md"
if not journal_path.exists():
raise HTTPException(status_code=404, detail="Journal not found")
entries = load_journal(journal_path)
exact = [e for e in entries if e.experiment_name == req.target]
if not exact:
raise HTTPException(
status_code=404,
detail=f"No journal entry for '{req.target}'. Run a backtest and record results first (fithia2 rec).",
)
entry = sorted(exact, key=lambda e: e.timestamp)[-1]
target_name = entry.experiment_name
def _only_this(e: Any) -> bool:
return e.experiment_name == target_name
with journal_lock(journal_path):
updated_count = refresh_public_scores(journal_path, selector=_only_this)
rebuild_registry(journal_path, registry_path, leaderboard_path)
fresh_entries = load_journal(journal_path)
fresh_exact = [e for e in fresh_entries if e.experiment_name == target_name]
fresh_entry = sorted(fresh_exact, key=lambda e: e.timestamp)[-1]
result = {
"experiment_name": target_name,
**_compute_for_entry(fresh_entry),
"entries_updated": updated_count,
}
return result