# Apples analysis (seed — for business context)

## Working on now

*This is a research notebook, not a finished product. It evolves as we work together.*

- **Validation tab is the v0.** It implements the Deflated Sharpe + Bonferroni-style required-t correction Ben raised on Apr 18. Pre-loaded with the 8 trial Sharpes from the existing trading project.
- **Open question for Ben:** what would you actually want a tool like this to do, beyond what's there? Hypothesis tracker? Per-window walk-forward harness? Live correlation monitor against canonical factors?
- **Next iteration after the call:** rebuild the walk-forward methodology in the trading project to re-optimize per training window, add Deflated Sharpe to the eval harness, redo the comparison with realistic Kraken fees (0.26% + ~0.2% slippage on breakouts).

---

## 1. The Email Thread (Oct 2024 – May 1, 2026)

Total direct exchanges: **7 messages**. The substantive thread starts April 18, 2026.

| # | Date | From | Substance |
|---|---|---|---|
| 1 | Oct 8, 2024 | Nik → Ben | Job-search ask. Sent Bitcoin price-prediction GitHub project + two resumes (IT + sales). Salutation "UB" (Uncle Ben). |
| 2 | Apr 18, 2026 | Ben → Nik | The big one. Recommends Barcelona as a base. Diagnoses Nik's bot as "**tantamount to data mining**". Walks through overfitting risk, walk-forward limits, and his preferred alternative: **start from an economic concept** (using ETF/NAV arbitrage as the canonical example). Recommends academic papers as the source for ideas. Closes with credentials (Booth, Steve, 10+ years) and a how's-life. |
| 3 | Apr 22, 2026 | Nik → Ben | Acknowledges the critique. Pivots to: building an AI consultancy at red.apples.name. Names the BTC halving cycle as a possible economic concept for crypto. Says bot is paper-trading, has built a research → strategy → results website. |
| 4 | Apr 29, 2026 | Nik → Ben | Updated apples.live link. Explains the per-client model: AI research + simple website + chatbot + then API integrations across Meta/Google/QuickBooks/Stripe/Shopify → automation. Sends Tener's preview link as example. |
| 5 | May 1, 2026 | Ben → Nik | "Great endeavor. Timing is perfect." Three direct questions: **Working solo? Any clients yet (Deda Bob)? Measurable impact?** |
| 6 | May 1, 2026 | Nik → Ben | Long answer: solo, no API connections yet from anyone, the value is in cross-platform pattern detection. Concrete examples: moped Facebook posts 3×/day, Canvas API for school deadlines, 180,000-business Google Business Profile email pipeline (free tier 100/day), Enformion + pool-permit follow-up flow for Bob. Offers Ben a personal preview link. |
| 7 | May 1, 2026 | Nik → Ben | "I think it would be cool if we set up a trading system this way, I'll send you a link once it's ready." |

**Status as of right now:** Nik just promised Ben a link (this analysis page is what fulfills that promise). The next ball in Ben's court is whether to actually click and engage.

---

## 2. The Trading-Project State Ben Was Reacting To

What Ben saw on his screen was a screenshot of the bot/dashboard. Here is the actual ground truth from `/root/projects/trading/` as of today, including the audit Nik already ran on himself after Ben's email.

### 2.1 The system

- **Pre-market scanner** (Yahoo Finance, 18 symbols, scored on trend/volume/momentum), AI commentary via Claude CLI, daily HTML briefings emailed and Slack/Telegram-pinged. Cron 8am ET (crypto) / 9am ET Mon-Fri (stocks).
- **Backtest engine** (`engine.py`) with strategy plugin interface, walk-forward harness, transaction-cost modeling, equity tracking.
- **Strategy library** (`strategies.py`): MA Crossover, Bollinger mean reversion, **Donchian Breakout** (declared winner), RSI mean reversion, Pivot Point mean reversion (legacy 15m, "no edge").
- **Live bot** (`bot.py`): currently paper-trading Donchian on Kraken, daily candles, per-asset parameters: BTC 70d/10d/2.5×ATR, ETH 55d/20d/3.0×ATR, SOL 30d/10d/1.5×ATR, XRP 30d/10d/2.5×ATR. 1.5% risk/trade. Systemd `trading-bot --dry-run`.
- **Dashboard** (`dashboard.py`, port 4066): strategy hypothesis + rules, signal proximity, equity curve, trade history, backtest comparison, performance footer.
- **Knowledge tree** (`learn.html`): expandable concept nodes tied to actual strategy.
- **Daily health check** (`/opt/trading-health.py`, 9pm ET): per-pair stats, RSI/levels, AI summary, HTML report emailed.

### 2.2 The self-audit Nik already produced (`ben_email_analysis.md`)

This file already exists and is *exactly* the kind of honest post-mortem Ben was pushing for. The seven red flags it surfaces:

| # | Issue | Evidence (file:line) |
|---|---|---|
| 1 | Params picked on full in-sample data, then labeled walk-forward | `optimize.py:41-79` |
| 2 | Walk-forward does **not** re-optimize per split; uses fixed params | `engine.py:454-456` |
| 3 | Per-asset custom params multiplies degrees of freedom | `CLAUDE.md` |
| 4 | 5 strategies tested, 1 winner kept, no Bonferroni / Deflated Sharpe correction | `CLAUDE.md:99-139` |
| 5 | Survivorship bias — BTC/ETH/SOL/XRP all survived; no LUNA/FTT/dead coins | asset selection |
| 6 | 26-35 trades per asset over 5 years is statistically thin (~6/yr) | `engine.py:319-332` |
| 7 | Fees too low: 0.1% commission + 0.05% slippage assumed; Kraken taker is **0.26%** + breakout slippage 0.15-0.25% → understated round-trip by ~40% | `engine.py:68-69` |

**Ben's diagnosis is fully validated.** The "Donchian works on all 4 assets" headline is in-sample optimization on a survivor-cherry-picked universe with thin trade counts and unrealistic fees.

---

## 3. The Two Frameworks Ben Is Pointing At

### 3.1 Overfitting / Probability of Backtest Overfitting (PBO)

Ben's professional language. The work to know:

- **Bailey, Borwein, López de Prado, Zhu (2014, 2016) — "Probability of Backtest Overfitting"** (Journal of Computational Finance). Defines PBO via combinatorially symmetric cross-validation (CSCV). If you tested N strategies and reported the best, this gives you the probability the live performance will be worse than median.
- **Bailey & López de Prado (2014) — "The Deflated Sharpe Ratio"**. Corrects Sharpe for selection bias under multiple testing **and** for non-normal returns. The "deflation" depends on number of trials, variance of trial Sharpes, skew, kurtosis. Crypto fails *all* the standard Sharpe assumptions, which makes DSR particularly necessary, not optional.
- **Harvey, Liu, Zhu (2016) — "…and the Cross-Section of Expected Returns"**. 316 published factors. Their bottom line: required t-stat for a new factor is **~3.0**, not 2.0, after multiple-testing. Roughly half of published anomalies don't survive.
- **Harvey & Liu (2014) — "Evaluating Trading Strategies"**. Sharpe haircut framework — given how many strategies you tested, what % of your reported Sharpe should you assume is real?

**Operating rule of thumb:** if you tested N strategies, required t-stat scales with √log(N). 100 strategies → t ≈ 3.4. Reported Sharpes typically need a 40–60% haircut.

### 3.2 Economic-concept strategies (Ben's preferred path)

The principle: the trade exists because there is a real-world mechanism that *forces* the deviation to resolve. Not "the chart did this last time."

| Concept | Mechanism | Retail viability today |
|---|---|---|
| **ETF / NAV arbitrage** (Ben's example) | Authorized Participants create/redeem in-kind, dragging ETF price toward NAV by close | Mostly **no** — AP-only. Residual retail plays: closed-end fund discounts, stressed bond-ETF discounts. Crypto analogue (GBTC/ETHE premium-discount) is **dead** as of Jan 2024 spot-BTC-ETF approval. |
| **Time-series momentum** (Moskowitz / Ooi / Pedersen 2012) | Long if asset's own 12m return is positive, short if negative, across 58 futures | Implementable retail with discipline; survives across asset classes |
| **Carry** (Koijen / Moskowitz / Pedersen / Vrugt 2018) | Long high-yield, short low-yield within an asset class | Implementable; for crypto the analogue is funding-rate carry |
| **Funding-rate arbitrage** (crypto-native) | Perp-vs-spot delta-neutral; capture funding when longs pay shorts (or vice versa) | **Yes, retail-accessible.** Reported 10-30% APY with sub-2% drawdown in delta-neutral setups. Risks: funding-flip, exchange counterparty, slippage on rebalancing, fees eating thin spreads. |
| **Basis trade** (CME futures vs spot) | Arbitrage CME futures premium/discount vs spot | Institutional-friendlier (margin, capital efficiency); retail-doable on smaller perps |
| **Stablecoin depeg reversion** | USDC/USDT depeg events resolve over hours-days due to redemption arbitrage | Tail-risk strategy; Mar 2023 USDC depeg was a textbook setup; rare events |
| **Cross-exchange arb** | Same asset trading at different prices on Coinbase vs Kraken vs Binance | Latency-sensitive, fee-sensitive, mostly squeezed by bots |
| **Bitcoin halving cycle** (Nik's suggestion) | Every ~4 years mining reward halves → supply shock | **Weakest** of the list. Three observations of an N=3 cycle is statistically inadequate; the "cycle" thesis is also debated. Use as long-horizon context, not as a tradable signal. |

**The retail-realistic short list for someone Nik's size ($500-5k starting capital):**
1. **Funding-rate arbitrage** — best risk/reward of the economically-grounded strategies that retail can actually run today.
2. **Time-series momentum at the daily/weekly horizon** — if you must do trend-following, do the academically validated form, not Donchian-on-4-cherry-picked-coins.
3. Avoid the rest until capital and infrastructure justify it.

---

## 4. Apples × Ben — The Integration Opportunity

This is the actual product question. What does Apples *do* for someone like Ben?

Ben is not a typical Apples client. He's not a service business that needs lead intake or review responses. He runs a **systematic trading firm** with a partner, an algorithm, and a fund. The Apples value proposition has to map differently.

### 4.1 Where Apples can add real value to Alturas / Ben specifically

**a. Research workflow automation.**
Ben said it himself: he reads countless academic papers, and Steve is a voracious markets reader. That's continuous ingestion + synthesis work that an AI system is genuinely good at. Apples can:

- Subscribe to **arXiv q-fin** + **SSRN finance** + selected journals; auto-summarize new papers daily; flag any that match Alturas's interest tags (S&P internals, regime detection, volatility, multi-strategy diversification).
- Maintain a Nikipedia-style internal wiki of strategies they've considered or are tracking.
- Build a "**concept-to-test pipeline**": when a paper is interesting, generate a one-pager, list the data needed, and queue a backtest stub.

**b. Backtest validation infrastructure.**
The thing Ben spent his entire April 18 email arguing for. Apples can host a private **PBO + Deflated Sharpe service**:

- Drop in trial Sharpes from any strategy variant, get back the deflated Sharpe + PBO + required t-stat for the trial count.
- Tracks every backtest ever run across the firm to maintain an honest "trials done" denominator (the most-violated rule in quant).
- Walk-forward harness with per-window re-optimization, realistic fees, multi-asset hold-out.

**c. Strategy diversification monitor.**
Ben said "any strategy can stop working at anytime, so we continue to diversify." Apples can:

- Track rolling correlation between Alturas's live strategy and a basket of canonical factors (TSMOM, carry, value, BAB, low-vol).
- Alert when realized correlation spikes — a sign the strategy's edge has migrated into a known factor and may be about to fade.
- Alert on PnL regime change (rolling Sharpe collapsing, drawdown duration extending).

**d. Operational automation across his existing tools.**
Even systematic firms have ops overhead. Apples connects:

- **Plaid / bank** → cash-balance tracking, fund accounting reconciliation.
- **Email (Gmail)** → LP/investor email triage, quarterly-letter drafting from performance data.
- **Calendar** → AI receptionist for investor meetings.
- **Drive / Sheets** → auto-update marketing decks with latest performance numbers.
- **GitHub** → pull internal repo activity into a weekly engineering digest.
- **Slack** → daily dashboard digest, alert routing.

**e. Communication / IR layer.**
Alturas Capital Castle Fund has LPs. Apples can:

- Draft monthly/quarterly performance commentary from the algo's own logs.
- Build an LP portal with auto-refreshing performance stats.
- Maintain a knowledge base of every question any LP has ever asked, with Ben-voice answers.

**f. Intellectual sparring partner.**
The thing Ben already does naturally with Steve. An LLM with full context of Alturas's strategy library, research notes, and trading history is a genuine third voice in strategy discussions — not replacing Steve, but accelerating the loop.

### 4.2 Concrete first-week ask

If Ben opts in to the preview link, the highest-leverage first integration is:

1. **Connect Gmail (read-only) + Calendar.** Lets the system see the existing investor / partner / vendor cadence and identify the top 3-5 recurring workflow points.
2. **Connect a single Drive folder** with sanitized strategy notes / academic-paper PDFs. Apples builds a private wiki, surfaces overlaps and gaps.
3. **One demo of the PBO/DSR service** running on the Donchian backtest from `/root/projects/trading/` as a worked example — Ben gets to see his own April 18 critique implemented as a tool.

This makes the value visible without asking him to expose Alturas's actual fund infrastructure. If demo (3) lands, the next step is offering it as a private deployment Ben can run on his own data, not on Apples-hosted infra.

## 5. Open Questions For Ben (when timing is right)

1. Is the Castle Fund strategy he runs at Alturas closer to **mean-reversion** on S&P internals, **trend** on S&P internals, or **regime classifier**? The SEC filing language ("price, volume, internal market data tied to one-day changes in the S&P 500 Index") is ambiguous.
2. How does he and Steve decide when to **retire** vs **tune** a strategy that is fading? (The exact question Nik raised in the unsent draft reply.)
3. What's a "good" first economic-concept strategy for someone starting at $500-5k of capital, in his view?
4. Would he consider a **paid or unpaid advisory role** on the trading project specifically? Even one-call-a-quarter has high value.
5. Does he have any interest in the AI-research-pipeline side of Apples for Alturas's own use? (Ingest arXiv/SSRN, summarize, surface relevance.)

---

*File maintained by Apples. Next refresh on any new Ben↔Nik email or any change in `/root/projects/trading/ben_email_analysis.md`.*

_To be enriched for oranges. Seeded from apples client `ben-rogers`._

_Pending — define what content this person would crave on oranges. Apples context lives in `interests` field above for reference._

{"subreddits": ["quant", "algotrading", "SecurityAnalysis", "investing", "finance", "econometrics", "AskEconomics", "financialindependence", "MachineLearning", "spain"], "youtube_channels": ["Patrick Boyle", "Quantopian", "CFA Institute", "Macro Alf", "Real Vision Finance"], "rss_feeds": ["https://hnrss.org/frontpage", "https://feeds.ssrn.com/ssrn-rss/SSRN_abstracts_by_date", "http://export.arxiv.org/rss/q-fin", "https://feeds.ft.com/ft/rss/home", "https://www.bloomberg.com/markets/rss", "https://www.risk.net/rss", "https://feeds.aps.org/rss/recent/physics.html", "https://marginalrevolution.com/feed"], "hn_keywords": ["quantitative finance", "backtesting overfitting", "systematic trading", "machine learning finance", "factor investing", "regime detection", "algorithmic trading", "academic finance"], "wikipedia_topics": ["Algorithmic trading", "Factor investing", "Modern portfolio theory", "Efficient-market hypothesis"], "humor_tone": "dry", "content_length_pref": "long", "language": "en", "novelty_appetite": "medium", "hard_avoid": ["crypto speculation", "get-rich-quick trading", "technical analysis astrology", "celebrity gossip", "sports"], "summary": "Academic-minded systematic trader who values economic-concept-grounded strategies, rigorous statistical validation, and long-form intellectual content across quant finance, macroeconomics, and European life.", "youtube_channel_ids": ["UCOIi9OLSpXRsX-wp2u26fsA", "UC606MUq45P3zFLa4VGKbxsg", "UCHwkW7lTTzf19xg5R4gfwVw", "UCFk1qCySNf2FIzIidVVW81A", "UCBH5VZE_Y4F3CMcPIzPEB5A"]}