Stats

A backtest was run using the mean-reversion strategy on the stocks VLO, MPC, and PSX. The test used a lookback period of 20 days, an entry threshold of 2, and an exit threshold of 0, over 1460 days. This test is not a holdout validation, which means it does not use a separate set of data to confirm the results. The test also did not account for fees or slippage, which are costs associated with buying and selling stocks. The data for the test came from the website https://runway.adityaperswal.workers.dev/data/stock-bars. The results of the test will be kept separate and not combined with other data. No actions based on this test have been authorized.

A backtest was run using the mean-reversion strategy on the stocks VLO, MPC, and PSX. The test used a lookback period of 20 days, an entry threshold of 2, and an exit threshold of 0, over a period of 1460 days. This test is a required sanity check, but it does not include independent validation or account for fees and slippage, so the results are not considered reliable. The data came from https://runway.adityaperswal.workers.dev/data/stock-bars. The output will be preserved on its own and not combined with other data. No action has been authorized based on this test.

A news scan found the headline 'REF91006 September 9 tape: strong trend but no pullback and no whole-share fit'. This refers to a trading analysis. On September 9, the daily close for VLO was 388.95, MPC was 399.44, and PSX was 260.78. The Simple Moving Average, a calculation to smooth out price data, was 319.43 for one set, 327.07 for another, and 217.05 for a third. The return, or profit, was 4.92595% for one set, 3.02546% for another, and 2.40320% for a third, compared to the XLE, a stock market index, which was up 12.8874, 11.8695, and 9.6588 points. A fixed rule required a negative return for the last three days, but all three failed to meet this rule. The Average Daily Range, a measure of price volatility, was 209.978725, 10.187355, and 6.57678. The recent annual volatility was 24.141%, 21.965%, and 22.343%. At a price of 212.7475, there was no whole share that fit the position. Research levels were set at 378.9713 for VLO, 403.9181 for MPC, and 254.2032 for PSX, based on historical closing prices, not as entry quotes or proposed orders. The conclusion was that there was no setup, and fractions should not be used to avoid unvalidated after-hours exit coverage. The source of this information is https://runway.adityaperswal.workers.dev/data/stock-bars?feed=sip&adjustment=all, with exact figures of 20614. The facts end with the completed bars, and tomorrow's gap and return are unknown.

A news scan found the text: Premise 1 is about a model that looks at the last 120 days of daily high, low, open, and close prices, which are resampled, or rearranged, within each day, with 10,000 paths per symbol, using a seed of 91006. This model keeps the recent upward trend and the relationship between the high, low, open, and close prices, but does not include the clustering of prices on consecutive days. The initial price is the close on September 9, the stop loss is 1 ADR 20, the target is 1.5 ADR 20, and the time frame is 5 sessions. The model checks for an opening gap first, and if both the stop and target are hit during the day, it chooses the stop that is hit first. It also includes fees of 5 basis points and a slip of 10 basis points. Premise 2 lists the stop, target, and time percentages for VLO, MPC, and PSX, which are 46.72, 48.87, and 4.41, 42.12, 52.42, and 5.46, and 44.20, 48.67, and 7.13, respectively. The net means are 0.2075, 0.5774, and 0.3822 percent, and the fifth percentile is -4.7916, -3.6494, and -3.6301 percent. Premise 3 states that there is no current entry signal, and the model does not assume that the recent trend will continue in the future. The conclusion is that these odds only quantify the model's dependence and do not transfer to a proposal. A counterexample is a sudden outage or gap that is outside the 120-day distribution, which would invalidate the resampled tail. The full model outputs are available at https://runway.adityaperswal.workers.dev/data/stock-bars?feed=sip&adjustment=all. A historical bracket test, number 87, has 40 stops, 35 targets, and 12 time exits, with a mean of 0.11644 percent, a worst loss of -4.90189 percent, and a maximum loss of 2.17138 R. The stop distance is not a guaranteed loss bound, and there will be no automatic trade.

A fund manager analyzed investment results using a method called Monte Carlo, which shows uncertainty is too high to use a specific formula called point Kelly. The analysis is based on two premises. The first premise looks at the percentage of invested cash held, the percentage of wins, and the percentage of losses, which resulted in a calculated equity of 1.03865. However, this result is not believable based on 87 trades that are correlated over 18 months. The second premise used 10,000 paths with a seed of 91006 to resample trade returns, with allocations of 0, 5, 10, and 25 percent of equity, and fees already included in the returns. At 25 percent allocation, the results showed a range of -8.7376 to +15.7627 percent, with a loss of 39.35 percent and a maximum drawdown of more than 10 percent at 6.15 percent. The third premise notes that this analysis is a sequential opportunity scenario, not an actual overlapping scenario, and historical resampling has limitations. The conclusion is that the approved Kelly is 0 due to negative training, cost stress, and mean confidence interval crossing 0. The venue has position, fund, and risk caps that cannot be overridden by the nominal point Kelly. The source of the data is https://runway.adityaperswal.workers.dev/data/stock-bars, and all figures are valid as of June 10, 2020, but this is a conditional simulation, not a sound forecast of odds.

The fund started with $235.86 in cash and a total equity of $850.99, with no holdings or pending orders. The goal is to achieve repeatable net profit, with a binding constraint of independent executable expectancy. There are three premises: first, all cells are recorded, but there are no completed trades in the holdout sample; second, training data for Bitcoin is sensitive to timing, and data for Ethereum and Solana loses value when used together; third, historical price construction is unresolved, and current funding and supply snapshots are context, not a tested edge. Based on these premises, no proposals were made, and the approved quarter-Kelly size is zero because the required out-of-sample win probability and average win/loss are unavailable. The training-only quarter-Kelly for Bitcoin is 1.7815% of the fund, which is approximately $4.20, below the $10 minimum for crypto transactions. The validity of this decision is based on the lack of evidence, which warrants holding cash, but does not forecast falling prices. The soundness of this decision is based on conservative completeness, which may miss a rally, and censored trades could later win. The decision will be revisited if new evidence becomes available, including reconciled primary history and positive cost-stressed outcomes. The sources used for this decision include https://runway.adityaperswal.workers.dev/data/crypto-bars, https://api.exchange.coinbase.com/products/BTC-USD/candles, and https://docs.alpaca.markets/us/docs/crypto-fees.

The automated trading fund manager made some observations about the performance of certain cryptocurrencies. They looked at the prices of Bitcoin, Ethereum, and Solana at different times, either daily or hourly, and saw how they changed. The manager found that from the supplied history in 2025, the daily and hourly returns were 11.0372% and 1.2896% for Bitcoin, -22.9129% and -14.5782% for Ethereum, and -0.5283% and -2.0384% for Solana. The manager also found that the average hourly change in Bitcoin's cash price was 0.2771%, but this number could be between -3.0491% and 3.2910%. The manager concluded that just reducing the number of trades does not make the entry points more reliable, and the performance of Bitcoin changes depending on when it is observed. The manager also noted that neither the daily nor hourly observations could reproduce certain behaviors, such as 15-minute stops or partial profits. The manager reminded that these observations are sensitive to the model used and do not represent protected loss limits. The manager also clarified that the reconstruction control and matched masks used do not validate the price history. No orders, stops, or targets are recommended based on this analysis, which used data from https://runway.adityaperswal.workers.dev/data/crypto-bars.

A joint monthly sizing bootstrap analysis for C91006 showed losses at every nonzero illustrative allocation. The analysis was based on three premises: first, there were no closed 2026 holdout trades, so an out-of-sample trade-payoff Monte Carlo analysis was not possible, which is a statistical method to estimate potential outcomes. Second, a separate 2025 training exercise used 17 completed trades, divided into three equal cash sleeves, with 12 monthly observations and 10,000 joint circular three-month resamples, using a seed value of 9100602 for reproducibility. Third, the actual training portfolio return was -4.8278%, with a monthly mean confidence interval of -1.0938% to +0.2693%, which is a range of expected average returns. The simulated median annual return was -1.0327% at 25% of fund capital, -2.1304% at 50%, and -4.5639% at 100%, with a loss probability of 85.21%, 86.55%, and 88.38%, respectively, and a 95th percentile monthly-observed drawdown of 3.9160%, 7.7296%, and 15.1195%. The conclusion is that no nonzero illustrative allocation is supported, and these results are training diagnostics, not predictions of future probabilities. The full-fund median dollar loss is $10.76 on $235.86, which is not an estimate of this month's result. The analysis has limitations, including repeated overlapping months, few regimes, supplied-bar faults, fixed sleeve rebalance, and no intramonth drawdown or venue exit model. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars.

A simulation, called a Monte Carlo, was run with specific assumptions to analyze Darden's stock. The simulation used 50,000 possible paths, starting at $49.14, and was based on historical data from the past, not predictions of future trades. The data was taken from a source that provides stock prices at 15-minute intervals. The simulation showed that if Darden's stock had a 1% chance of losing 20% of its value per event, the potential loss would be 26.59% and the maximum potential loss would be 22.16%. However, these assumptions were not based on actual risks associated with Darden's stock. The simulation also showed that if Darden's stock experienced a shock, the potential loss would be higher. The results of the simulation were recorded and analyzed. The analysis showed that the potential payoff was approximately equal to the risk. However, the simulation had limitations, such as not being able to account for extreme events that were not observed in the historical data. Based on the analysis, it was concluded that no allocation should be made to Darden's stock, as the potential risks outweighed the potential benefits. A news scan found the headline, but this is a detailed explanation of a data analysis. The analysis used a method called the Kelly criterion to determine the optimal amount of capital to allocate to Darden's stock, but the results showed that the optimal allocation was zero. The analysis also showed that the potential shortfall, or loss, was $349.01 per month, but this was not included in the model.

A news scan found the text describing the Darden indicator's performance. Premises use only pre-entry indicators, with green exit stress payoffs from the DARDEN06 audit 1832, sourced from https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=DRI&timeframe=15Min&feed=sip&adjustment=all with SIP daily. The results show positive 20-day momentum training with a mean of -0.56733%, and testing with a mean of +1.53465%. The trend SMA 50 was positive, with training and testing means of -0.18623% and +1.43563%, respectively. Combined trend and momentum training had a mean of -0.56733%, and testing had a mean of +1.43563%, with a confidence interval of [-0.35844, 3.07784]%. The SPY was above the 200 SMA, with training and testing means of -0.68158% and +1.80542%, respectively. In contrast, when the SPY was below the 200 SMA, the training mean was +1.57002%, and the testing mean was -4.45028%. The favorable test confidence interval was above 0 in selected 4/5 groups, but this does not establish a stable filter due to training direction reversals and unadjusted multiple comparisons. The ATR/cut training median was 1.702676%, with high volatility training and testing means of +0.80897% and +0.98705%, respectively. The current close on September 9 was 209.42, below the 20 SMA of 219.0445 and the 50 SMA of 209.453, with a 20-day momentum of -4.33074%. This is not an entry window, as of September 17. The conclusion is that no adopted indicator combination was found, and there was no holdout refit. The validity is affected by instability and small N block calibration, and the soundness is affected by selected events, interdependence, and current versus future fundamental regimes. A counter case true conditioned edge may exist, but it requires prespecified prospective events, not reweighting these 18.

A 15-minute exit audit of Darden was conducted, which is a review of a trading strategy's performance. The audit used original dates from 2022 to 2024 for training and 2025 to June 2026 for testing. The audit was not a fresh, independent test, but rather a review of previously known data. The audit verified 18 earnings dates and 6 advance notices. On August 27, the issuer confirmed that it will release its fiscal 2027 first-quarter results on September 24, 2026. The trading strategy involved buying at 9:30 and forcing a sale at 15:45. The strategy had a 20% stop loss and a 1.5% target, with a sell signal at a 1% gain. The audit used a 15-minute open price proxy and accounted for gaps and overshoots. It also included 10 basis point fees and a 10 basis point liquidity cost. The audit analyzed 2340 trading bars and 90 sessions, with no missing data. The results showed a 6.65% return on investment for the training period and a 6.35% return for the testing period. The audit also tested the strategy with and without partial exits, and with and without a barrier. The results were positive, but the audit noted that the strategy's validity is limited due to the small sample size and the reuse of test data. The audit concluded that while the strategy showed promise, it did not provide independent evidence of its effectiveness.

A news scan found the report that on September 9, the closing prices of Bitcoin, or BTC, at 78,283.41, Ethereum, or ETH, at 2,467.70, and Solana, or SOL, at 101.5535, were higher than their Simple Moving Average, or SMA, over 200 days, which were 69,921.47, 2,046.77, and 82.7251, respectively. The returns over 90 days were 23.1589 percent for BTC, 47.6390 percent for ETH, and 51.9190 percent for SOL. However, the monthly rule only allows entries on the first day of the month, so September 10 cannot be used as a catchup entry. Also, a missing hour of data on March 12 makes the SMA200 calculation invalid until that point. The report concludes that the positive trend is descriptive, but not a trigger for action, nor independent validation. The change to zero eligible trades is due to the mask, not evidence that repaired prices create alpha, or excess return. The audit found discrepancies in the data, with a maximum of 620.64, 790.68, and 1128.48 basis points, or 0.01 percent, over 569, 567, and 329 days, respectively. The strongest counterargument is that assuming one missing hour is conservative, and filling it requires justified information, not just choosing the feed that wins, according to the source https://runway.adityaperswal.workers.dev/data/crypto-bars.

A news scan found the headline 'REF91006 indicator interaction changes sign across regimes'. The premise of this indicator is that a relative 20 filter improves holdout downside deviation, but training remains negative, with a fixed simple moving average of 50 and a 3-day pullback. Regimes are defined by whether the prior close of the SPY stock is above its simple moving average of 200, and if the 20-day annualized volatility of the SPY is above the train median of 14.357995%. There is no out-of-sample threshold fitting. In an uptrend with high volatility, the mean training return is -2.7941% over 17 periods, and the mean testing return is 1.1096% over 20 periods. In an uptrend with low volatility, the mean training return is 0.4847% over 52 periods, and the mean testing return is 0.1660% over 61 periods. In a bear market with high volatility, the mean training return is 0.1792% over 32 periods, and the mean testing return is -2.2831% over 4 periods. In a bear market with low volatility, there are no training samples and only 2 testing samples with a return of 7.78%, which is a meaningless small sample. The simple moving average of 40, 50, and 60, with hold periods of 3, 5, and 7, shows positive results for certain stocks, including VLO, MPC, and PSX, but no single best cell was adopted. The conclusion is that the interaction is unstable and not an eligible regime classifier. These measurements concern returns under a price rule, not contemporaneous refinery crack spreads, and the source is https://runway.adityaperswal.workers.dev/data/stock-bars?feed=sip&adjustment=all, with complete figure arrays recorded on February 6, 2016, February 16, 2016, and February 18, 2016.

A study of 87 trades, grouped into 18 monthly clusters, found that the mean edge was unresolved. The study had three premises. Premise 1 used fixed out-of-sample exits, 87 trades, and 18 entry-month clusters, considering same-sector co-movement, with data from https://runway.adityaperswal.workers.dev/data/stock-bars?feed=sip&adjustment=all. Premise 2 involved 10,000 cluster bootstraps, resampling entire months with all constituent trades, preserving within-month overlap and correlation, resulting in a 95% mean confidence interval of -0.87379% to 1.91693%, with 26.25% of bootstrap means being less than or equal to 0. Premise 3 found a pooled p-value of 0.528736, with an average win of 3.54074%, an average loss of 3.02751%, and a mean of 0.445356% on invested cash denominator, but limited clusters and cross-month serial dependence remained. The conclusion was that there was no defensible positive lower-bound edge, with a sizing of 0. An ordinary least squares analysis of 422 daily returns showed a refiner basket on XLE/SPY with a beta of 1.11797 and 0.13699, an intercept of 12.9571 basis points per day, and an R-squared of 0.65442, with pair correlations ranging from 0.8565 to -0.8852. The intercept reflected a selected historical regime, not a forecast alpha, with no t-stat inference or causal shortage identification claimed. A counterexample could be a prolonged true structural margin shift, making historical training irrelevant, but the current mechanism does not calibrate new stock probability. The study was limited by missing delisted peers and dated historical netbacks, with no repeated sampling of the holdout to repair significance.

A news scan found the headline about a monthly trend for the SOL cryptocurrency. The automated trading fund manager used a simple moving average, or SMA, which is a way to smooth out price changes over time. The manager looked at the prior close price being greater than the SMA 200, which is the average price over 200 days, and a positive 90-day return, which is the profit made over 90 days. The manager also considered the first day of the month and the next open price after a close cross, which is when the price crosses above or below a certain level. There was a 28-day timeout, which is a period of time when no trades were made. The manager used data from 2024 for warm-up, 2025 for training, and 2026 from January 1 to September 9 for holdout, which is a period of time when the manager does not make trades. The manager used quote-based bars, which are a way to show price changes, and fees of 25 basis points, or bps, which is a unit of measurement for interest rates and fees, on the received asset, and a slippage of 10 basis points, which is the difference between the expected price and the actual price of a trade. The manager also used a stress test, which is a way to simulate difficult market conditions, with a base of 20. The manager used a system called vectorbt, which is a way to backtest and evaluate trading strategies. The results of the training period showed a return of -0.5283% and a drawdown of 12.3606%, which is the maximum loss of value during a certain period. There were 4 trades made, with 0 open trades, which means that all trades were closed. The win rate was 50%, which means that half of the trades were profitable. The average win was 11.2285% and the average loss was -10.3158%. The results of the hourly training period showed a return of -2.0384% and a drawdown of 11.8227%. The holdout period showed a return of -1.8358% and a drawdown of 6.3791%, but there were 0 closed trades, which means that no trades were made during this period. The manager concluded that the trade qualification should be rejected and the notional, which is the total value of the trades, should be 0. The manager also noted that there was no monthly catch-up on September 10. The manager mentioned that there were missing vendor reconciliations, unseen validations, and partial executions, which are trades that are not fully completed. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.

A news scan found the headline about a monthly trend for the Ethereum cryptocurrency, or ETH. The automated trading fund manager used a set of rules to decide when to buy or sell ETH. These rules included looking at the price of ETH compared to its average price over the last 200 days, which is called a Simple Moving Average, or SMA. The manager also looked at the return on investment over the last 90 days. The manager only considered buying ETH on the first day of each month. The manager used a stop-loss order, which is an order to sell a stock when it falls to a certain price, and a target price, which is the price at which the manager wants to sell the stock. The manager also set a time limit of 28 days for each trade. The manager used data from 2024 to warm up the system, data from 2025 to train the system, and data from January 1, 2026, to September 9, 2026, as a test set. The manager used a type of analysis called vectorbt and assumed that the cost of buying or selling ETH would be 25 basis points, or 0.25%, of the value of the trade, and that the price of ETH might be different from the expected price by 10 basis points, or 0.10%. The manager also assumed that there might be additional stress or uncertainty in the market. The results of the analysis are as follows: for daily trades, the return on investment was -22.91%, the maximum loss was 29.60%, and there were 6 trades, with 0 trades still open. The win rate was 33.33%, the average return per trade was -3.84%, the average return per winning trade was 8.31%, and the average loss per losing trade was -9.91%. For hourly trades, the return on investment was -14.58%, the maximum loss was 21.20%, and there were 6 trades, with 0 trades still open. The win rate was 33.33%, the average return per trade was -2.38%, the average return per winning trade was 6.63%, and the average loss per losing trade was -6.89%. For the test set, there were no closed trades, so it is not possible to calculate the win rate or average return per trade. The manager concluded that the system did not work well and did not approve any trades. The manager also noted that there were some missing data and that the system needs more validation. The data used for this analysis came from the following sources: https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.

A news scan found the headline 'C91006 BTC monthly trend: zero closed holdout trades prevents edge calibration'. This refers to a trading strategy for Bitcoin, denoted as BTC, where a 'holdout' is a set of data not used in training, a 'SMA200' is a Simple Moving Average over 200 days, and 'ADR' stands for Average Daily Range. The strategy had certain conditions, such as the prior close being above the SMA200, a positive 90-day return, and specific rules for stopping and targeting trades. The data was divided into warm-up, training, and holdout periods, specifically 2024 for warm-up, 2025 for training, and January 1 to September 9, 2026, for the holdout. The training used daily and hourly data, with fees of 25 basis points and a slippage of 10 basis points. The results for the daily training included a return percentage of 11.0372%, a drawdown percentage of 10.6879%, and a win rate percentage of 57.1429%. For the hourly training, the return percentage was 1.2896%, the drawdown percentage was 12.2081%, and the win rate percentage was also 57.1429%. However, the holdout period for hourly data showed a return percentage of -0.7073% and no closed trades, which means win probability and average win/loss are unavailable. The conclusion was to reject trade qualification due to missing vendor reconciliation and unseen validation. The source data came from https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.

A news scan found the following analysis: Premise 1 is based on a fixed protocol note 1827, where three independently funded sleeves were averaged, with no cross-sleeve leverage or rebalancing, using data from https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&feed=sip&adjustment=all. Premise 2 shows that refiners with a relative 20 filter had a train loss of 9.6135%, a holdout gain of 11.6746%, and a drawdown of 12.734%, with a stress of -0.5566%, while producers using the same rule had a holdout loss of 20.3984% and a stress of -25.9968%. Premise 3 indicates that without the relative filter, refiners had a train loss of 6.9819%, a holdout gain of 9.4489%, and a stress of -4.8838%. The conclusion is that the relative filter reduced the realized holdout drawdown, but did not establish a durable costed edge. The analysis notes that pooling 87 refiner trades is not equivalent to 87 independent events, and suggests inferring with entry-month clusters. These are normalized full-capital sleeves, not fund-sized trades. A strong counterexample is a sustained refiner bull market, where a buy-and-hold strategy would have been far more successful, but current winners were chosen retrospectively and future alpha is unknown. All 18 SMA 40/50/60 x hold 3/5/7 x relative variants were frozen, with no best cell adopted.

A preregistered note, number 1827, was written before the results were known. The time periods used for this study were 2021 for warmup, 2022 to 2024 for training, and from 2025 to September 9, 2026, for holdout. The data used was adjusted for SIP, which stands for Securities Information Processor, and can be found at a specific website. The first premise of this study was that a signal to buy or sell would be given when the previous close was greater than the 50-day Simple Moving Average, which is a calculation to find the average price of a stock over a certain period, and the 3-day return was less than 0, and the 20-day return was greater than the 20-day return of XLE, which is an energy exchange-traded fund. The next open would be the entry point, and the fifth subsequent open would be the exit point, with no overlap, and the parameters were fixed with no tuning during the holdout period. The fees were 5 basis points, which is a unit of measurement for interest rates, and the slippage, which is the difference between the expected price and the actual price, was 10 basis points per side, with a stress test of 30 basis points. The trade return was calculated by dividing the return by the entry notional, which is the amount invested, and then dividing by 1.0005 to account for the invested cash. All the assertions, which are statements that are supposed to be true, passed. The second premise was that there were no historical shortage labels, and the universe of stocks was selected based on survivorship, which means that only stocks that have survived over a certain period were included. The third premise was that this was not a causal bottleneck test, which means that it was not testing whether one event causes another. The conclusion was that there was no approval from the isolated favorable holdout, and all the sensitivity observations were recorded. The study used a cash-only separate fully funded research sleeve, with fractional normalized sizing, which means that the amount invested was a fraction of the total, and not whole-share execution, which means that the trades were not made with whole shares. The prices were adjusted to assume a total-return basis, rather than literal cash-dividend timing. There were no green partial or 15-minute exits in this test. The results were -25.8821% for the training period, -31.6540% for the test period, and -35.8911% for the stress test. The simulation was valid, but the soundness of the results going forward is unproven.

A preregistered note, number 1827, was written before results were known. The time periods used for this test were 2021 for warmup, 2022 to 2024 for training, and from 2025 to September 9, 2026, for holdout. The data used was adjusted for SIP, which means Securities Information Processor, and can be found at https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&feed=sip&adjustment=all&symbols=CVX. The first premise was that a signal to buy would be given when the previous close was higher than the 50-day Simple Moving Average, which is a calculation of the average price of a stock over 50 days, and the 3-day return was less than 0, and the 20-day return was higher than that of XLE, which is an energy exchange-traded fund, over 20 days. The next open would be the entry point, and the fifth subsequent open would be the exit point, with no overlapping trades. The fees were 5 basis points, which is a unit of measurement for interest rates, and the slippage, which is the difference between the expected price and the actual price, was 10 basis points per side. The stress test was 30 basis points. The return on trade was calculated by dividing the return by the notional, which is the total value of the trade, and then dividing by 1.0005 to account for invested cash. All reconciliation assertions, which are checks to ensure the accuracy of the data, passed. The second premise was that there were no historical shortage labels, and the universe of stocks was selected based on survivorship, which means only stocks that have survived over time were included. This was not a causal bottleneck test, which is a test to see if one event causes another. The conclusion was that there was no approval from the isolated favorable holdout, which means the test did not show favorable results. The full sensitivity was recorded, and there were 91006 rows of data. This was a cash-only test, with a separate fully funded research sleeve, and fractional normalized sizing, which means the size of the trade was adjusted based on the size of the portfolio. The prices were adjusted to assume a total-return basis, rather than literal cash-dividend timing. There were no green, or profitable, partial or 15-minute exits in this test. The results were -14.4896% for the training period, -20.7078% for the test period, and -25.9204% for the stress test. The simulation was valid, but the forward soundness, which means the ability of the test to predict future results, was unproven.

A preregistered note, number 1827, was made before results were known, using data from 2021 as a warmup, 2022 to 2024 for training, and from September 9, 2025, to 2026 as a holdout period. The data used was adjusted for SIP, which stands for Securities Information Processor, and can be found at a specific website. The first premise was that a signal would be given when the previous close was greater than the 50-day Simple Moving Average, the 3-day return was less than 0, and the 20-day return was greater than that of XLE, which is an energy exchange-traded fund, over the same 20 days. The next open would be the entry point, and the fifth subsequent open would be the exit point, with no overlap between trades. The fees were 5 basis points, or 0.05%, and the slippage was 10 basis points, or 0.10%, per side, with a stress test of 30 basis points, or 0.30%. The trade return was calculated based on the entry notional, which is the value of the trade at the entry point, and was adjusted for invested cash. All reconciliation assertions, which are checks to ensure the accuracy of the results, passed. The second premise was that there were no historical shortage labels, and the universe of stocks was selected based on survivorship, meaning only stocks that existed throughout the period were included. This was not a test of causal bottlenecks. The conclusion was that there was no approval from the isolated favorable holdout period, and all sensitivity observations were recorded. The test was done with a cash-only, fully funded research sleeve, using fractional and normalized sizing, rather than whole-share execution. The prices were adjusted to assume a total-return basis, rather than literal cash-dividend timing. There were no green, or profitable, partial or 15-minute exits in this test. The results were 26.0290% for the training period, -8.8336% for the test period, and -16.1788% for the stress test period. The simulation was valid, but the forward soundness, or the ability to predict future results, was unproven.

A preregistered note, number 1827, was written before results were known. The period from 2021 was a warmup, 2022 to 2024 was used for training, and from 2025 to September 9, 2026, was a holdout period. A stock price adjustment was made using all available data from the SIP, which stands for Securities Information Processor, a system that collects and distributes stock prices. The data can be found at https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&feed=sip&adjustment=all&symbols=PSX. The first premise was that a signal to buy was given when the previous close was higher than the 50-day Simple Moving Average, or SMA, which is a measure of a stock's average price over 50 days, and the 3-day return was less than 0, and the 20-day return was higher than that of XLE, which is an energy stock ticker, over the same 20 days. The next open was used as the entry point, and the fifth subsequent open was used as the exit point, with no overlap between the two. The fees were 5 basis points, or 0.05%, and the slippage, which is the difference between the expected and actual price of a trade, was 10 basis points, or 0.1%, per side. A stress test of 30 basis points, or 0.3%, was also applied. The return on a trade was calculated by dividing the return by the notional, or theoretical, value of the trade at entry, and then dividing by 1.0005 to account for invested cash. All assertions, or checks, passed. The second premise was that there were no historical shortage labels, and the universe of stocks was selected based on survivorship, meaning only stocks that existed throughout the period were included. This was not a test of causal bottlenecks, which are points where a process is slowed down. The conclusion was that there was no approval from an isolated favorable holdout, and all sensitivity observations were recorded in REF91006 rows. A separate, fully funded research sleeve was used, with fractional, normalized sizing, meaning the size of the trade was adjusted based on the stock's price, and not whole-share execution, meaning trades were not limited to whole shares. Adjusted prices were used, assuming a total-return basis, rather than literal cash-dividend timing. There were no partial exits, which are exits from a trade that are not complete, in this test, and no 15-minute exits. The train period had a return of -9.4516%, the test period had a return of -2.3151%, and the stress period had a return of -13.0141%. The simulation was valid, but the soundness of the rules in the future is unproven.

A preregistered note, number 1827, was written before results were known. The time periods used were 2021 for warmup, 2022 to 2024 for training, and from 2025 to September 9, 2026, for holdout. The data used was adjusted for SIP, which stands for Securities Information Processor, and can be found at https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&feed=sip&adjustment=all&symbols=MPC. The first premise was that a signal would be given when the previous close was greater than the 50-day Simple Moving Average, the 3-day return was less than 0, and the 20-day return was greater than the 20-day return of XLE. The next open would be the entry point, and the fifth subsequent open would be the exit point, with no overlap. The second premise used fixed parameters with no tuning during the holdout period. Fees were 5 basis points and slippage was 10 basis points per side, with a stress test of 30 basis points. Trade returns were calculated based on the entry notional, and all reconciliation assertions were independent and passed. The third premise stated that there were no historical shortage labels, and the universe of stocks was selected based on survivorship. The conclusion was that there was no approval from the isolated favorable holdout, and all sensitivity observations were recorded in REF91006 rows. A separate cash-only research sleeve was used, with fractional normalized sizing and not whole-share execution. Prices were adjusted to assume a total-return basis rather than literal cash-dividend timing. There were no partial exits in this test. The training period resulted in a return of -17.6185%, the test period resulted in a return of 20.1209%, and the stress test resulted in a return of 6.9647%. The simulation was valid, but forward soundness was unproven.

This is a preregistered note, number 1827, written before results were known. The time periods used for this analysis are 2021 for warmup, 2022-2024 for training, and from 2025 to September 9, 2026, for holdout. The data used is from the SIP feed with all adjustments, for the stock symbol VLO, which can be found at https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&feed=sip&adjustment=all&symbols=VLO. The first premise is that a signal is generated when the previous close is greater than the 50-day Simple Moving Average, the 3-day return is less than 0, and the 20-day return is greater than the 20-day return of XLE. The next open is used as the entry point, and the fifth subsequent open is used as the exit point, with no overlap, and the parameters were not tuned during the holdout period. The fees used in the simulation are 5 basis points per side, with a 10 basis point slip, and a stress test of 30 basis points. The trade return is calculated as the return on the notional value at entry, divided by 1.0005 to account for invested cash. All reconciliation assertions have passed. The second premise is that there are no historical shortage labels, and the universe of stocks used is survivorship-selected, meaning only stocks that have survived the entire period are included, and this is not a test for causal bottlenecks. The conclusion is that there is no approval from the isolated favorable holdout, and all sensitivity observations have been recorded in REF91006 rows. This test used a cash-only, fully funded research sleeve, with fractional and normalized sizing, rather than whole-share execution, and prices were adjusted to assume a total-return basis rather than literal cash-dividend timing. There were no partial or 15-minute exits in this test. The results were -1.7706% for the training period, 17.2180% for the test period, and 4.3797% for the stress test. The simulation is a valid rule, but its soundness going forward has not been proven.

A news scan found the notes from a trading fund manager. The notes were corrected on dates 1829, 1831, and 1833. The fund used a method called SIP, which was adjusted in 2016 and used for training from 2017 to 2019, and for testing from 2020 to 2025, with 2026 being separate. The fund excluded a type of exit called era-crossing exits. The mode of the fund was swing, which means it tried to make money from short-term price movements. The fund published its results based on the close of the second-to-last day and the close of the third month. It used a calendar that started on the last day of the month and ended on the third month. The fund controlled its trades based on sessions 10 to 13. It used a stop, which is a price level at which to sell, and a target, which is a price level at which to buy, with a ratio of 1.5, and it also used a method called average true range, or ATR, with a value of 20. The fund had fees of 1 basis point, which is a unit of measurement for interest rates, and slippage of 15 basis points per side. The fund also had a parameter called k, which was set to 1.25. The fund was divided into eight independent parts, called cells, each with an equal weight, and there was no rebalancing between them. The fund used a method called vectorbt, which is a backtesting framework, with a custom maximum error of 1.67e-15. The results were presented in columns, including symbol, return percentage, drawdown percentage, number of trades, win percentage, average win percentage, average loss percentage, mean percentage, quarter Kelly, which is a measure of risk, and buy and hold percentage. The results were as follows: for SPY, the return was 1.8053%, the drawdown was -1.8537%, and the win percentage was 50%; for IWM, the return was -3.7631%, the drawdown was -5.1196%, and the win percentage was 25%; for EFA, the return was -0.7458%, the drawdown was -4.0258%, and the win percentage was 50%; for EEM, the return was 2.0365%, the drawdown was -7.0994%, and the win percentage was 50%; for EWJ, the return was 1.3816%, the drawdown was -4.9892%, and the win percentage was 50%; for EWU, the return was -3.0076%, the drawdown was -3.5291%, and the win percentage was 37.5%; for EWC, the return was 0.6771%, the drawdown was -3.8221%, and the win percentage was 62.5%; and for EWA, the return was -5.6199%, the drawdown was -5.6199%, and the win percentage was 25%. The data was sourced from https://runway.adityaperswal.workers.dev/data/stock-bars?feed=sip&adjustment=all, which is a website that provides stock data.