CLAUDE.md

@AGENTS.md

The target: alpha vs QQQ, not return

Judge, rank, promote and blend bots by their annualised alpha against QQQ and its t-stat. Do not use raw return, and do not use Sharpe on its own. This applies whenever you evaluate a bot, pick C2 / copier weights, or decide whether a new strategy is worth building.

Why: QQQ exposure costs nothing, because anyone can buy QQQ. A bot only earns its place if it adds return that QQQ does not already provide. Raw return rewards the wrong things:

• In a rally it rewards beta. A beta-1 bot "beats" everything and is just QQQ with extra moving parts. A leveraged bot looks even better until the drawdown. For example, RecursiveDecayHarvestBot had beta 2.6 and alpha −81%/yr, yet matched QQQ on return.
• It punishes low-beta bots for doing their job. A beta-0.25 bot in a +17% tape should make about 4%. That is not underperformance.
• Only uncorrelated alpha improves a portfolio. A high-correlation bot adds nothing to a QQQ holding. A low-correlation positive-alpha bot raises the Sharpe of the whole book. Since beta can be bought directly, alpha is the only scarce ingredient.

How to compute it: use portfolio_worth daily series and follow the weekday/gap rules in the AGENTS.md "PortfolioWorth Model" section. Drop weekend rows, and drop day-pairs more than 4 days apart (recorder outages). Then:

beta = cov(r_bot, r_qqq) / var(r_qqq)
resid = r_bot - beta * r_qqq
alpha = resid.mean() * 252  # annualised
t_stat = resid.mean() / resid.std() * sqrt(n)  # n = daily obs

Report alpha, t-stat, beta, correlation and max drawdown together, each over the bot's own live window, with QQQ over that same window.

The weekly alphareport CronJob does exactly this for every live bot (Saturdays 07:00 UTC, utils/alpha_report.py). It writes the bot_alpha_report table with a verdict per bot:

• edge;
• pause candidate;
• levered QQQ / QQQ clone;
• unproven;
• too short.

Start from the latest report_date there rather than recomputing by hand. backtest_bot uses the same alpha_stats function.

Backtests use the same method. backtest_bot returns alpha, alpha_t, beta and benchmark_corr (see _compute_alpha_metrics in tradingbot/utils/backtest.py), and local_optimize / tune_hyperparameters rank by alpha_t by default. Don't tune a bot on yearly_return or sharpe_ratio, because that selects for beta. When you pass data= yourself, also pass benchmark_close=. Otherwise the alpha keys are None, and that means "not measured", not zero.

Idle cash earns T-bill yield on the broker side. The live copier parks uninvested weight in SHV (LIVETRADE_CASH_PROXY; see LiveTradeCopier._park_idle_cash). Paper bots still hold 0%-yield USD, so their portfolio_worth alpha understates what the copied book earns by roughly rf × the cash share.

The bar:

• Real edge requires t ≥ 2. Below that it is noise, however good it looks. At 75–120 live days (typical as of Sep 2026) almost nothing clears it, so say "unproven" rather than "working".
• Prefer beta < 0.5 / correlation < 0.5. A bot with beta ≈ 1 and ~0 alpha is a QQQ clone. Replace it with QQQ, don't blend it.
• Negative alpha with t ≤ −2 is a real result: pause the bot.
• Prefer a lower-alpha, lower-correlation bot over a higher-return, higher-beta one.

Snapshot, 2026-09-21 (live, per-bot window; none reached t ≥ 2 positive):

Bot	Alpha/yr	t	Beta	Verdict
TARegimeAdaptiveBot	+14%	1.30	0.31	likely luck: 6y backtest alpha +0.4% (t 0.10)
XAUZenbotTreeBot	+20%	0.88	0.05	dead: flat on every bar since ^XAU > 325, not refittable
AdaptiveMeanReversionBot	~0%	0.19	0.99	QQQ clone
FearGreedBotQQQInverse	+5%	0.56	0.91	QQQ clone
RecursiveDecayHarvestBot	−81%	−1.93	2.61	levered QQQ, bleeds — paused
EURUSDTreeBot	−11%	−2.32	0.02	significantly negative — paused

The C2 blend at the time (Kronos / RegimeAdaptive / EarningsInsiderTilt) made +5.5% vs QQQ +17.4%, with alpha +3% at t = 0.25. That is weak on the metric that matters.

Follow-ups from that snapshot: SqueezeMomentumBot and StockNewsSentimentBot are also paused (suspend: true in values.yaml). FearGreedBotQQQInverse is now flat by default (buy at ≤30, exit at ≥50). TARegimeMultiAssetBot runs as paper only. Both of those reliably cut beta by about half or more in backtests, but neither has alpha that holds up in both halves of the sample. See docs/backtests/taregimemultiassetbot.md and the comment in feargreedbot.py.

Second pass (same day):

• Also paused: SynthesizedHyperConvexityBot (−26%/yr alpha, 3x TQQQ/SQQQ), AIHedgeFundBot and DeepSeekToolBot. The two AI bots had beta ~0.7 with negative alpha, and every run makes paid LLM calls. aiHedgeFund.suspend pauses the AIHedgeFundBot feeder too.
• TRADEABLE has no leveraged or inverse ETFs. RegimeAdaptive, EarningsInsiderTilt and SharpePortfolioOpt sell TQQQ/UPRO/FAS/TMF/SQQQ on their next rebalance.
• Re-tunes are walk-forward. Grid-search on 2019–2022, then judge the winner against the defaults on 2023–2026. New parameters ship only if they beat the defaults out of sample. For the harness, see docs/backtests/retune-2026-09.md.

Third pass (2026-09-22):

• The C2 blend is now TARegimeMultiAsset 0.39 / GoldenButterfly 0.37 / TSMOM 0.24, inverse-vol weights fit on the first half only. Backtest alpha +3.5%/yr at t 2.02, positive in both halves, beta 0.13. It is the first candidate to clear t ≥ 2. See docs/backtests/c2-blend-2026-09.md.
• The live no-trade band is at stage 2 ($25 / 5%). That matches the code defaults, so live execution and backtests now use the same band.
• An earnings-drift (PEAD) sleeve was tested and not built. On 93 large caps it does not beat holding the same universe. See docs/backtests/pead-2026-09.md.
• Insider data was broken until this date. yfinance's Transaction column is blank, so every stock_insider_trades row was untyped and EarningsTilt's insider score was always 0. Any analysis of that bot before 2026-09-22 reflects earnings surprise only.