Nancy Pelosi Stock Trading Bot¶
This bot estimates Nancy Pelosi’s publicly disclosed household stock and listed-partnership portfolio, then scales that basket to your account. It reads official House reports to reconstruct the full inventory. Delayed disclosures and value ranges make this an approximation of her holdings.
Want her call options too? The Nancy Pelosi Copy Trading Bot copies those as well.
How it works¶
The bot has just two AI agents, and both run every day:
Research agent reads the newest eligible annual snapshot and subsequent transactions on the House Clerk website. It returns every included issuer’s numeric asset-value range, snapshot date, eligible share changes and sources. Transactions after the snapshot date apply once, even if filed before the annual report. Full sales remove a holding; partial sales reduce it.
Trading agent estimates shares from each range midpoint and its snapshot-date price, applies later share changes and values the remaining inventory at current prices. It allocates 99% of account capital across the complete feasible basket. Current quotes size trades; small constituents remain part of replication. New disclosures trigger a full rebalance, including exits and additions. On quiet days established positions within 2 percentage points of target stay in place.
There is no other code. Want a different member of Congress? Change the name in the research agent’s prompt.
Run it on BotSpot¶
Run this bot on BotSpot without installing anything. BotSpot runs LumiBot in the cloud, backtests it, and connects it to your broker.
The code¶
The whole bot is one short file. The two prompts are the strategy. The research agent uses LumiBot’s built-in web tools, which read ZIP files, PDFs, spreadsheets and web pages from any website.
"""Estimate and mirror the publicly disclosed household stock portfolio.
Disclosures are delayed and use value ranges, so this is an approximate basket.
"""
from lumibot.strategies import Strategy
class NancyPelosiTradingBot(Strategy):
def initialize(self):
self.sleeptime = "1D"
self.agents.create(
name="researcher", model="openai/gpt-6-luna", reasoning_effort="high",
allow_trading=False, allow_network=True, model_request_timeout_seconds=0, run_timeout_seconds=0,
system_prompt=(
"Reconstruct Nancy Pelosi's full publicly disclosed household listed-stock and listed-partnership "
"portfolio as of as_of. This is an estimate, not verification of her exact account. Find the "
"official yearly indexes at https://disclosures-clerk.house.gov/public_disc/financial-pdfs/<year>FD.zip. "
"Check the as_of year and earlier years needed for the baseline. Indexed transaction PDFs use "
"https://disclosures-clerk.house.gov/public_disc/ptr-pdfs/<filing-year>/<DocID>.pdf; annual reports "
"use financial-pdfs/<report-year>/<DocID>.pdf under the same root. Use actual indexed IDs. "
"Start with the newest eligible annual snapshot; use its reporting year-end when implicit and "
"disclose that convention. Apply eligible transactions AFTER that valuation date once, including "
"transactions filed before the annual. Earlier transactions are already reflected in the baseline. "
"Never use a disclosure known to be after as_of. Read every asset-section text chunk, including "
"wrapped issuer names, tickers, type labels and value ranges. Return the actual numeric lower and "
"upper ASSET VALUE bounds for every included annual issuer, with report ID and page or text chunk. "
"None ASSET VALUE is zero modeled baseline exposure; None INCOME does not erase a positive value. "
"Audit names against the full numeric ledger and separate zero exclusions. Resolve discrepancies "
"rather than inventing a holding or dropping a wrapped row. Include later stock and listed-partnership "
"additions: unit purchases count regardless of bracket type labels, as do shares from exercised calls. "
"Treat each security row separately: an issuer can have both share and option purchases. Classify "
"each transaction from its own description and share or contract quantity, not another row for that issuer. "
"Exclude unexercised options, funds, bonds, private assets and property. Distinguish full sales from partial sales: an explicit full exit removes that "
"holding, not merely a share subtraction from an uncertain baseline estimate. Partial sales, "
"gifts and contributions reduce shares; exercises add shares. Reconcile amendments, splits, "
"spinoffs, mergers and delistings from eligible evidence. Do not apply an estimated spinoff again "
"when a later filing confirms it. Return BOTH snapshot and as_of tickers for every issuer. "
"Resolve renamed listings from dated issuer announcements; a stopped ticker is not an exited issuer. For Block, only on/after 2025-01-09 check its dated ticker announcement at "
"https://investors.block.xyz/investor-news/news-details/2025/Block-Announces-Ticker-Symbol-Change-to-XYZ-To-Report-Fourth-Quarter-Results/default.aspx; apply changes only when effective by as_of. "
"Return compact JSON with as_of, snapshot_date, ledger, zero_exclusions, source_checks and limitations. "
"The ledger's issuers are the union of positive annual stock/unit rows and later share additions, "
"including new issuers created by exercises, purchases or spinoffs. Each issuer row gives "
"snapshot_ticker, as_of_ticker, baseline_lower, baseline_upper, baseline_kind, share_changes and full_exit. "
"Each share_changes entry gives its transaction_date, signed numeric shares, kind and source_id. "
"A new issuer has a modeled zero baseline and its known positive share changes; include it even "
"when the annual lists only options. An exercise label without the numeric share count is incomplete. "
"Keep all positive additions in this one ledger. Include applied IDs and latest filing checked. "
"Check new filings every iteration. Without new evidence retain the full prior ledger with "
"refreshed source status; a lookup failure is not a sale. Label stale or approximate evidence. The trader handles valuation and sizing. Do not trade."
),
)
self.agents.create(
name="trader", model="openai/gpt-6-luna", reasoning_effort="high",
allow_trading=True, model_request_timeout_seconds=0, run_timeout_seconds=0,
system_prompt=(
"Mirror the full source-ledger stock and listed-partnership basket, scaled to this account. "
"Estimate unknown baseline shares from the asset-value midpoint and the snapshot-date price, "
"apply eligible later share changes once, and mark remaining shares at current prices. An "
"explicit full sale sets the holding to zero; do not invent a residual from midpoint uncertainty. "
"Recheck contradictory quantities and dates. Use a labeled baseline-value proxy only when "
"quantity reconciliation or an eligible historical price genuinely remains unavailable. Check "
"issuer, listing, USD currency and consistent split/share basis; use historical ticker aliases "
"when needed. Normalize estimated remaining values into portfolio weights. Target 99% of current account capital with a joint cash-aware allocation. Buy the complete "
"feasible basket initially, without waiting for a new disclosure. Small positions are part of "
"replication, not discretionary conviction bets. For a positive allocation smaller than one "
"share, prefer one whole share using unspent joint cash above the reserve first; trim larger "
"allocations only if funding requires it and disclose rounding. Never short, overspend or use margin. "
"Reconcile every positive source name against targets and actual holdings. On quiet days leave "
"established holdings within 2 percentage points of target alone, after all feasible constituents "
"are held. Complete missing holdings with funding trims when needed. New eligible disclosures "
"require rebalancing the updated whole basket, including additions and full exits even when "
"individual drift is small. Use the researched as_of listing for current quotes and snapshot ticker for old prices. "
"For a still-unpriced issuer, use its baseline value midpoint as a labeled denominator proxy and reserve "
"that estimated allocation in cash; complete the supported remainder. Preserve supported holdings if research fails; nothing new "
"is not a liquidation signal. Report complete target versus actual weights, fills, cash and specific unresolved exceptions."
),
)
def on_trading_iteration(self):
as_of = self.get_datetime().isoformat()
previous = self.vars.get("last_research", "none yet")
research = self.agents["researcher"].run(
task_prompt="Return the complete numeric source ledger with snapshot and as_of tickers, inventory audit, eligible changes and newly checked filings.",
context={"as_of": as_of, "your_last_answer": previous},
)
self.vars.set("last_research", research.summary)
self.agents["trader"].run(
task_prompt="Value the full disclosed basket and complete every feasible target position change in this account.",
context={"as_of": as_of, "research": research.summary, "previous_research": previous},
)
if __name__ == "__main__":
from lumibot.credentials import IS_BACKTESTING
if IS_BACKTESTING:
from lumibot.backtesting import YahooDataBacktesting
NancyPelosiTradingBot.backtest(YahooDataBacktesting)
else:
NancyPelosiTradingBot().run_live()
Run it yourself¶
pip install lumibot ddgs==9.16.1 primp==2.0.1
python -m lumibot.example_strategies.ai_nancy_pelosi_trading_bot
The two extra packages supply the existing keyless web-search tool. Direct document reads remain available; installing the search backend gives the researcher another way to find eligible original reports.
The local example selects Luna High with API funding. Configure OPENAI_API_KEY and your broker keys (for example ALPACA_API_KEY, ALPACA_API_SECRET and ALPACA_IS_PAPER=true). For subscription funding, follow Use your ChatGPT subscription and select chatgpt/gpt-6-luna for both agents. Hosted BotSpot runs use BotSpot AI credits.
With IS_BACKTESTING=false the bot trades. With IS_BACKTESTING=true it backtests; set BACKTESTING_START and BACKTESTING_END to choose the period, and BACKTESTING_BUDGET to choose starting cash. The default starting account is $100,000. For Polygon history, set BACKTESTING_DATA_SOURCE=polygon and configure your own POLYGON_API_KEY with access to the required dates. Both agents run every iteration, so a longer test makes more model calls.
Good to know¶
Members of Congress can take up to 45 days to report a trade, so the bot always sees trades late.
Reports show dollar ranges, such as $1,000,001 to $5,000,000. Midpoints, snapshot prices and later disclosed changes produce estimated weights, not verified account balances.
The yearly report comes out in May and shows what she owned on December 31. Until the new one is filed, the bot starts from the year before and adds every trade since.
In a backtest the website also shows reports from after the test date. The research agent reads the date on each report and skips any report filed after the test day.
Whole-share rounding can leave cash or slightly overrepresent a small constituent. A missing price is a specific exception; the supported remainder still gets allocated. A source lookup failure does not mean an existing holding was sold.
Pelosi said on November 6, 2025 that she will not run for re-election in 2026 (NBC News). Her reports stop after she leaves office, so change the name in the research prompt to keep the bot trading.
The House website says its reports may not be used for most commercial purposes (5 U.S.C. 13107). Check that your use is allowed.
See AI Trading Bot Examples for more AI trading bots and Backtest, paper, or live: choose the runner for backtest and live runs.