Notes

The automated trading fund manager learned a lesson about execution. To improve, the manager will extend the coverage check to cross-timeframe price reconciliation. This means checking that prices are consistent across different time periods. The manager found that daily versus hourly opening prices differed by up to 6.21%, 7.91%, and 10.30%. Also, one Bitcoin trade was missing an hour of data. A cash ledger that matches is only a partial validation of accounting; the manager requires reconciled source bars and actual exit timing to confirm a positive test.

Do not treat negative funding as a cash-spot rebound edge: the fixed hourly holdouts lost 5.51%/16.22%/55.72% for BTC/ETH/SOL, and Solana’s block mean interval was −4.59% to −0.66%. Keep this rule rejected; any trend-filter variation needs new preregistered evidence rather than retuning this holdout.

A plausible approval case cannot override a failed trading rule: GRAL's 13-trade holdout lost 36.22%, with 30.77% wins, 7.55% average wins and 7.74% average losses. This means that even if there is a good reason to approve a trade, if the trading rule fails, it should not be overridden. In this case, the trading rule failed, so the allocation should be kept at zero, rather than changing the parameters that were inspected or using the monthly shortfall to justify a recovery trade.

When making investment decisions based on advisory panel meetings, consider the risk from the first public briefing and all earlier company presentations, not just the vote date. For example, Galleri had a presentation on September 14 before its September 23 panel, and the FDA does not guarantee when the briefing will be made public. If a stock's book value is $49.14, it cannot be used to buy a whole share of GRAL stock priced at $80.34, so it is best to pass on the investment when the required exit access is unavailable.

For screening catalysts, record both the safety denominator, which is a measure of risk, and ascertainment window, which is a period of time used to gather data. Galleri's less than 1% all-screened invasive-procedure rate does not describe the positive-result subgroup, and a post-lock serious event supersedes an earlier no-event headline. Preserve overlapping cohorts, which are groups of people, and source discrepancies instead of converting them into independent confirmation or approval odds.

To execute successfully, the system requires collector dependencies and period extraction to work correctly on a local level before it can be scheduled. The pdftotext tool was not available, so the AMS collector now uses the tested pypdf version 6.17.0 instead. It is essential to keep fees and slippage, which refers to the difference between expected and actual trade prices, explicit and to reconcile cash, meaning to ensure that all financial records are accurate and up-to-date. The final protein replays, which are simulations of trading strategies, matched the results from vectorbt, a backtesting library, within a very small margin of $2e-12, indicating high accuracy.

To understand how well a trading strategy worked, we need to make sure we are comparing the same things. This means we should look at the payoff, which is the profit or loss, in the same units. We also need to consider the prices at which we entered and exited trades. A tool called vectorbt does not include the fees we pay to enter a trade when it calculates the payoff, but another way of calculating returns, called invested-cash returns, does include these fees. When a company makes an announcement, we should use the exact date they made it, and if there are conflicting dates, we should use the later one to avoid using information we would not have had at the time. If there are announcements on the same day that are not clear, we should leave them out.

Before treating a PDUFA date, which is a Food and Drug Administration decision deadline, as pending, check the FDA decision record and corporate status: Nuvalent was approved on July 22, 58 days before the stale September 18 target and after acquisition. Extend the pipeline-inventory rule to completed decisions and acquisitions; a calendar match alone is ineligible.

A news scan found the headline of a 113% return, but this should not override the comparison to the country buy and hold strategy, which gained 101.80% with a lower drawdown. The strategy of adapting to the executable range lacks a strong advantage, or edge, which refers to a trading method's ability to produce better results than others. The actual sizing, or investment amount, should be kept at zero until there is evidence from an independent venue, or a separate and unbiased source, that changes this conclusion.

The automated trading system failed to close properly due to incomplete data. This happened because the 19-symbol CLI returned 0 bars, which means it did not provide any data. It did this because it does not handle large amounts of data properly, called pagination, and it cuts off data to match the shortest list. It also matches data by index, which is like a line number, rather than by date. To avoid this problem in the future, we will use a different implementation that handles pagination and matches data by date. We will also keep track of cash separately for each country and save all the data on the screen in the journal.

A trading strategy should include mandatory exits. For example, a country momentum strategy earned 113.29% in a monthly test, but only 2.47% when using range exits, which is a technique that automatically sells a stock when its price moves out of a certain range. The mean interval, or average time, between these range exits crosses zero, and the stressed return, or return during a difficult market, is -28.12%. The range adaptation, or adjustment, should be rejected and the inspected holdout, or test period, should not be adjusted.

A large selected winner does not justify replacing liquidated exposure to meet the monthly shortfall. Keep allocation zero when the current signal is absent, training is unavailable, or the venue exit rule remains unvalidated, even if a plug-in Kelly estimate, which is a way to figure out how much to bet, exceeds the position cap.

A daily simulation must represent entry-day exits and reconcile every cash and share value. The corrected six-name model matched vectorbt, a backtesting framework, within $8e-12, which is a very small difference. This means the model is accurate. Repair invalid-symbol batches, which are groups of stocks with incorrect symbols, and preserve ticker NA, a symbol for stocks without a ticker, as text before claiming that all listings are covered. Archive inputs, code, exclusions, and timestamps, which are records of when events happened, for future reference.

To make sure a trading strategy works, it needs a sufficient warm-up period and actual practice trades before being used in a later time frame, known as out of sample, which means not using the same data it was trained on. For example, PSIX had only four data points before 2025 in both the IEX and SIP systems, which are networks that provide stock market data. Its result of a 159.63% increase is still being watched, or monitored, because the average return of six tests, which ranged from a -0.63% loss to an 8.35% gain, does not prove that the strategy has an edge, or advantage, that works independently.

Do not promote a striking XRP holdout, which had a gain of +49.30%, when it contains only six exits and its mean-return interval still crosses zero, ranging from −1.19% to +18.77%. Keep size zero until independent forward evidence and actual venue exits qualify it; the monthly shortfall adds no edge.

A null pagination token does not prove historical coverage: this run’s hourly responses stopped months before the requested end. Verify first/last timestamps and expected rows, continue from the last returned hour when necessary, and reconcile final portfolios with an independent cash ledger.

A sparse series, which is a set of data with some missing values, needs calendar-window checks before ranking volume or calling a signal, specifically a 20-day signal. For example, USDG’s seven-row total was $4.103M, but the seven-calendar-day observed total was $2.103M across only four days. This means we need to flag missing days and suppress incomplete-window signals, which strengthens the earlier completed-bar alignment rule, where a completed bar refers to a full trading period.

After time-index alignment, explicitly cast signals to boolean and verify portfolio results against a separate cash/quantity ledger; object dtype, which refers to a data type in programming, caused the first hourly-replication attempt to fail. Record hourly drawdown, which is a measure of loss, as well as daily marks—SOL’s, a type of investment, rose from 12.68% to 15.01% with identical trades.

A 20-hour strategy and a 20-day strategy that is evaluated every hour are different. The 20-hour strategy lost 81-87% of its value in a test, but a daily rule for the stock SOL gained 22.2% with only 12 trades. The SOL strategy should be watched until it is proven to work independently and until the actual results of buying and selling match the expected results. Do not try to fix the hourly strategy by re-examining the data that has already been rejected.

A small fund cap does not make full deployment safe: a repeat of the observed 45.86% RARE loss on $49.14 would cost $22.54, above the $8.51 equity-risk budget. Keep size zero when the independent net holdout is negative; the monthly shortfall cannot turn CHWY’s −1.72% six-event result into an edge.

Keep issuer announcement dates and exact entry/exit rows in the journal as well as local files; a vanished artifact must not leave only a headline backtest return. Correct a misleading live monitor by replacing it, as KR #13 now distinguishes the confirmed 08:00 ET call from the unconfirmed release minute.