Bot Bought – Automated Paper-Trading System (Python, SQL, FinBERT)

Can a simple dip-buying rule beat the market – and how would I know?

Rather than hunting for a profitable backtest, I used this project to practice rigorous analysis. Every strategy claim had to survive four tests before being believed:

• Monte Carlo null baseline: is the result distinguishable from random entry timing?
• Risk-adjusted benchmark: how does it compare to SPY buy-and-hold on Sharpe ratio?
• A/B test: does a news-sentiment filter improve results?
• Power analysis: how much data would a real verdict need?

The Answer

The strategy does not beat SPY, and the sentiment filter hurt returns. That negative result is the point of the project: I let the data say no.

One engine, config-driven. Runs daily at 3:45 pm ET on weekdays on a headless Ubuntu server, scheduled with systemd timers.

1. 1 Data fetch Pulls daily bars for 16 stocks + SPY benchmark
2. 2 Signal Flags a dip: drop ≥ 1.5× the stock's 20-day ATR AND below its 20-day moving average; skips stocks within 3 days of earnings
3. 3 Sentiment FinBERT scores recent headlines; negative-news dips are vetoed (shadow mode: scored and logged, not enforced)
4. 4 Risk $500 fixed position, max 6 open positions, $400 max daily loss, kill switch, cash only
5. 5 Execute Paper orders via Alpaca; exits at +6% target, −6% stop, or 10 trading days
6. 6 Log Every signal, gate decision, veto (with the headlines behind it), order, fill, and reconciliation stored in SQLite
7. 7 Report Nightly text digest, weekly review package, automated backups

Trading Rules

• Entry: drop ≥ 1.5× the stock's 20-day ATR (volatility-normalized) AND below its 20-day moving average; skips stocks within 3 days of earnings
• Sentiment filter: FinBERT scores recent headlines; negative-news dips are vetoed. Currently in "shadow mode": scored and logged, not enforced (see Results for why)
• Risk guardrails: hard checks before any order: $500 fixed position, max 6 open positions, $400 max daily loss, kill switch, cash only. No automated process can change these.
• Exits: +6% target, −6% stop, or 10 trading days

Design Choices

• ATR-normalized entry instead of a flat % drop: "a 3% drop in KO is an event; in TSLA it's Tuesday"
• Avoided RSI<30 on purpose: the most-taught retail signal is the most crowded
• Watchlist chosen for sector spread (max 3/sector), varied volatility, and varied news coverage, so the sentiment filter could be tested where news is thick vs. thin
• $10k paper balance (not the $100k default) so results transfer to realistic account sizes

Project Facts

My role	Designed and directed, built with Claude Code. Sole owner: designed the strategy, set constraints, directed the build, and verified every result.
Timeline	Started Aug 11, 2026; live paper trading since mid-Aug 2026; ongoing
Status	Build complete (one optional test step left); in live forward-testing / review phase
Budget	$0 – free data, free APIs, open-source model, reused hardware
Money at risk	None – Alpaca paper account ($10,000 simulated)
Scale	~13,000 lines of Python source, 1,504 automated tests, 99% line coverage on the risk module
Backtest	10.6 years (2016–2026), 16 tickers, 45,000+ daily bars, 487 simulated round-trip trades

Tech Stack

• Languages/data: Python (pandas), SQL (SQLite – trade log, schema migrations, append-only audit tables)
• Data sources: Alpaca API (prices, news, order execution), yfinance (historical bars), FRED (macro)
• ML/NLP: FinBERT (HuggingFace transformers) – financial-news sentiment scoring, run locally on CPU
• Infrastructure: Ubuntu Server (headless), systemd timers for scheduling, SSH, Git/GitHub, nightly automated backups
• Testing: pytest-style unit tests, mutation testing (deliberately breaking code to confirm tests catch it)
• AI tooling: Claude Code (implementation), Claude (planning, review, analysis)

All results below are from the backtest (2016–2026, live parameters, nothing tuned). They are not live trading results.

Backtest vs. SPY Buy-and-Hold

Metric	Strategy	SPY buy-and-hold
Total return	+21.4%	+305.8%
Sharpe ratio	0.644 (best parameter cell)	0.839
Round-trip trades	487	–

The raw return comparison is unfair (the strategy averaged only 7.2% of capital deployed), so I used Sharpe ratio for the fair comparison. Buy-and-hold still wins on a risk-adjusted basis.

Null Baseline (Monte Carlo Permutation Test)

Measure	Value
Test design	100 seeds × 2 generators × 2 fill-cost models
Real strategy's percentile among random entry schedules	94th–97th
Pre-registered threshold for a 9-cell parameter grid (family-wise correction)	99.43rd
How often a cell hits the 97th percentile by chance	24% of the time
Verdict	NOT statistically distinguishable from chance

Sharper finding: when dates were kept but tickers shuffled, the edge nearly vanished. Timing carries what edge exists; stock selection doesn't.

Sentiment Filter A/B Test (2020–2026)

Metric	Filter ON	Filter OFF
EV per trade	+$2.42	+$3.68
Sharpe ratio	0.313	0.497
Total, ON vs. OFF	−$545
By year	ON worse in every year except 2023

Trades	Win rate
Blocked by the filter	58.5%
Kept by the filter	55.9%

• Mechanism: a dip big enough to trigger usually HAS negative news – that's what moved the price – so a negative-news filter removes the best rebounds.
• Checked whether this was just COVID: it wasn't. Blocked trades spread across every year.
• The filter is structurally a mega-cap filter: low-coverage stocks (KO, CAT) often have no news to read.

Market-Regime Analysis

Measure	Value
Share of trading days carrying 32% of all signals	6.8%
Signals on those days where the 6-position cap binds	64%

Signals cluster in market-wide shocks (Brexit, Feb 2018, COVID, April 2025 tariffs).

Exits: Retracting My Own Recommendation

56% of trades exit on the 10-day time limit. That initially looked like "wasted" trades, but timeouts were the LARGEST profit contributor:

Exit type	Profit contribution
10-day time limit	+$1,311
Target / stop	+$795

I retracted my own earlier recommendation to tighten exits. A 5-day max hold was clearly worse; 10 vs. 20 days was a plateau.

Win Rate vs. EV

The strategy with the best win rate (58.3%) earned less per trade than one winning 54.4%. Ranking by win rate inverted the right answer, so I chose EV per trade and Sharpe ratio as the metrics that matter.

Timing vs. Selection

Shuffling tickers while keeping dates nearly erased the edge: what edge exists comes from when the strategy buys, not which stock it picks.

Data Quality Catches

• Caught a stock-split bug producing false signals at 20–30× ATR; fixed via split-adjusted data
• Caught the backtester computing ATR differently from the live bot (87 of 136 rows disagreed) and made them identical
• Found the vendor revised "final" volume data after the fact; narrowed a validation rule that had been over-trusted and pre-registered a test of when bars become final
• Found a verification script reporting 163 "blocking" disagreements, 158 of which were a stale data snapshot

Engineering Rigor

• Pre-registration: predictions and decision thresholds written down before running tests, to avoid fitting a story after the fact
• Verify the verifier: planted known defects in comparison tools to confirm they detect them before trusting a "0 differences" result
• Mutation testing on the money path (order submission/recovery)
• Append-only audit trail: trade log and decision register can't be edited, only annotated

• Paper trading only. No real money is at risk, and no real profits or losses exist.
• Backtest results, not live results. The live forward record is small (1 trade open, 0 closed as of 9/17/26). The backtest numbers above are the substantive results; there is no live performance data yet.
• No out-of-sample split, by reasoned choice. 2020 held most stress events, so any chronological holdout would be either crisis-free or crisis-only – a holdout that can't fail returns "validated," which is worse than none. I used the null baseline instead.
• The sentiment filter can't be tested evenly. Low-coverage stocks (KO, CAT) often have no news to read, so the filter is structurally a mega-cap filter.

Power Analysis: How Much Data a Verdict Needs

Measure	Value
Per-trade standard deviation	$29.16
Confidence interval at the v1 target of 30 trades	±$10.44/trade
Backtest edge to detect	+$4.33/trade
Trades needed to detect it	~175 (~3.8 years at the measured rate)

So the review process was designed NOT to wait on a verdict that 30 trades can't deliver.

As of 2026-09-17.

• Live paper trading since mid-Aug 2026; 31 daily runs, 476 signals evaluated
• First strategy-chosen trade: BAC, Sep 16, 2026 (1.60× ATR dip, sentiment scored neutral → passed)
• First sentiment filter decision logged the same day
• v1 goal: 30+ closed trades, projected ~April 2027 at the measured signal rate
• Next milestone: final v1 report comparing live results to SPY (report script already built, waits on data)

Source code is currently private.