v1.0.0 — Titanium Alpha

1.0.0 -- 2026-04-20

First public release. All core components validated end-to-end on 15 years of
market data (2011-2026) and 1002 passing tests.

Added

• Walk-forward benchmark with two-cycle retrain (126 days) + rebalance (15
  days), volatility targeting at 10% annualised, transaction costs (5bps
  slippage + 10bps commission), and 16-metric portfolio vs SPY report.
  10-year OOS champion: Sharpe=0.766, CAGR=13.68%, MaxDD=-21.94%, Beta=0.566,
  Alpha=+2.57%.
• CPCV-OOS parameter tuning with Deflated Sharpe Ratio (Bailey & Lopez de
  Prado 2014). 982 configurations tested across three tiers; champion config
  committed in run_benchmark.py.
• PatchTST forecaster (NeuralForecast) with multi-quantile loss (5
  quantiles), CDF interpolation for continuous prob_up, CDF rearrangement
  for monotonic quantiles, and NaN/Inf guards in the predict path.
• LangGraph 4-agent debate (Technical Analyst, Fundamentalist, Bear,
  Portfolio Manager) with Pydantic-structured outputs and streaming per-node
  callbacks. Gemini (default, free tier) or Anthropic Claude via
  LLM_PROVIDER=anthropic.
• Financial RAG with sentence-transformers embeddings (all-MiniLM-L6-v2)
  and ChromaDB vector store; P95 retrieval latency = 101ms on 172 articles.
• HRP allocator with Ward linkage, Ledoit-Wolf covariance shrinkage,
  sum-preserving confidence tilt, and waterfilling weight optimiser.
• DecisionEngine with three-tier weight model (BUY=HRP, HOLD=HRP*conf,
  SELL=0); cash implicit; fallback to predictions.parquet if agents fail.
• Streamlit dashboard with four tabs: Benchmark, Performance, War Room
  (live + replay modes), Microstructure (fan chart).
• Data pipeline with thread-safe yf.Ticker().history() ingestion for
  52 US large caps + SPY, 15 years OHLCV (~199k rows), news ingestion with
  Google News RSS fallback.
• .claude/agents/ with 5 development subagents (architect,
  quant-reviewer, security-data, test-writer, docs-writer) encoding project
  rules and historical incidents.
• Docker Compose for PostgreSQL + ChromaDB.
• CI/CD via GitHub Actions (ci.yml + lint.yml).
• Methodology notebook (notebooks/methodology_and_results.ipynb)
  executing end-to-end in 6.6s with 9 sections covering problem framing,
  data, CPCV-OOS, DSR, HRP, walk-forward, regime analysis, the max_weight
  lesson, and limitations.

Fixed

• Thread-safety bug in yf.download() (pre-session 37): 22 of 52 tickers
  were receiving data from alphabetically-adjacent tickers due to a yfinance
  thread-safety defect. Migrated to yf.Ticker().history(); 15 years of data
  re-ingested.
• Look-ahead bias in early Sharpe estimates: the infamous Sharpe ~2.7 from
  sessions 1-35 was an artefact of look-ahead in rolling windows and fill
  strategies. Corrected to the current Sharpe=0.766 with full CPCV-OOS and
  out-of-sample walk-forward.
• Geometric rf conversion across benchmark_metrics, cpcv,
  cpcv_oos, and walk_forward (was arithmetic, biased Sharpe upward).
• Transaction costs silently dropping from port_ret in walk-forward
  (session 36).
• prob_up discretisation: replaced step function with CDF interpolation
  across 5 quantiles.
• Fan chart sort order: was alphabetical, now sorted by quantile level.
• Volatility targeting applied ex-ante (pre-allocation) instead of ex-post
  (post-allocation) to avoid leverage lag.
• HRP tilt sum-preserving: previously naive multiplication could drift
  the sum of weights away from 1.
• walk_forward initial state: portfolio starts 100% cash (institutional
  convention) rather than fully invested.
• Bankruptcy safeguard in walk-forward (guards against NaV<=0 chains).
• Dashboard (app.py:1396): decision.get('weight') fallback to
  suggested_weight for live debate mode (weight key only exists after HRP
  merge in replay mode).
• Streamlit deprecations: 17 use_container_width=True replaced with
  width='stretch' (deadline 2025-12-31).

Known limitations

• Backtest-production gap: the production make decide pipeline uses the
  LangGraph debate, which is not part of the walk-forward validation (cost
  of ~26k LLM calls per rebalance is infeasible). See
  docs/design_gap_backtest_vs_production.md.
• Stochastic debate: make decide is non-deterministic (temperature=0.2
  analysts + 0.1 PM). decisions.json is a single-run snapshot; borderline
  tickers flip between BUY/HOLD in reruns. The MIN_CONFIDENCE_FOR_ACTION=0.3
  gate is the deciding boundary.
• RAG coverage depends on news availability: with 172 articles across 52
  tickers, grounding coverage is ~47% on tickers with real debate. Meta and
  Tesla sparse in Google News RSS; citations will vary with the backfill.

Methodological stance

• The project favours honest reporting over marketing numbers. MaxDD,
  Beta, Alpha, Tracking Error are published alongside Sharpe; the
  backtest-production gap is documented in the README; the session 40
  regression (max_weight=0.10 reverted at Sharpe=0.462) is preserved as
  evidence of anti-overfitting discipline.
• max_weight is treated as a risk constraint (min(6%, 2/N)) not an
  optimisation parameter; fine-tuning via CPCV-OOS is valid only for timing
  parameters (rebalance_every, target_vol).