Stats

The Catalyst ledger was audited at 12:30. The results are as follows: the ledger contains 3 closed losses and 0 open or pending trades, with 2 operator liquidations that did not meet their intended horizons. A Python analysis of the ledger found the following returns: MIRM -0.439321%, IONS -2.842420%, and RARE -45.860062%, with a mean loss of 16.380601% and a win rate of 0/3. The average winner is undefined. The sample size is too small to determine if all future edges will be negative. The RARE exit price was 13.98976, which is 41.7093% worse than the stop price of 24. A separate analysis of CHWY stock from August 25 to September 8 found a negative observation, with an open price of 24.47 on August 25 and 23.86 on September 8, and a 5 basis point fee and slip on each side, resulting in a net loss of 2.687668%. This new observation was not added to the historical holdout data, which has the following parameters: N6, p50%, W4.523391%, L4.828695%, mean -0.152652%, return -1.719988%, and DD 14.258614%. The costs and corrected capital Kelly formula imply non-positive capital for CHWY, so the adopted capital is 0. The sources for the runway are query_trades?fund=catalyst and https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=CHWY,LOVE,SPY&timeframe=1Day&feed=iex&adjustment=all. Note that a basis point is a unit of measurement, in this case, a 5 basis point fee is 0.05%.

A news scan found the headline about Tyra Biosciences' Phase 2 results. The results show confidence intervals, which are ranges of values, for certain medical outcomes: for TYRA, the intervals are 28.86–82.34% for one outcome and 63.06–100% for another. These intervals assume independent outcomes, meaning each outcome is separate, and do not account for multiple tests or biases. Another set of results, for IRD, shows intervals of 19.41–99.37% and 47.82–100%. A query of trades, which are buys and sells of stocks, shows the following results: SNOW decreased by 5.39236%, DELL increased by 7.93994%, and GTLB decreased by 5.13050%. The average win, or gain, was 7.93994 and the average loss was 5.26143. All trades were liquidated, or sold, before the 4-6 week horizon, so the outcomes are incomplete for the original rule. The conclusion is that neither the clinical results nor the trade history show a positive edge, which is a trading advantage. A new study with exact endpoint classification, which is a way of measuring outcomes, would be needed to change this conclusion. Sources include query_trades fund, which is a type of automated trading fund, and press releases from Tyra Biosciences and Opus Genetics, which are companies. The calculations are archived under observation numbers 18398, 18400, and 18401.

A trading-run arithmetic check was performed, which is not a new backtest or new independent samples. The check reused recorded costed 2025 training and 2026 holdouts from 1639 to 1641 and a correction from 1653. The check was done using Python, which computed the average win and loss for each cryptocurrency. The formula used was p times the average win minus 1 minus p times the average loss, and the maximum of 0 and 0.25 times p minus 1 minus p divided by the win over loss. For Bitcoin, the probability of a win, p, was 42.6829%, the average win was 2.642605%, and the average loss was 3.080586%, with 82 trades, resulting in an expectancy of -0.637761%. For Ethereum, p was 44.2623%, the average win was 4.139850%, and the average loss was 4.236197%, with 61 trades, resulting in an expectancy of -0.528767%. For Solana, p was 47.1429%, the average win was 4.496455%, and the average loss was 4.301185%, with 70 trades, resulting in an expectancy of -0.153726%. All quarter-Kelly weights remained zero. The 15-minute compounded returns were -43.5409% for Bitcoin, -32.9637% for Ethereum, and -18.7330% for Solana, and the daily and hourly returns were also negative. These tests do not validate previously rejected pullback family rules, and the conditional holdouts only had 1 to 5 trades, with cost and training failures remaining frozen. The trade history had one forced liquidation, with no setup win or loss, or slippage distribution. The premises required a net positive demonstrated payoff, which was absent, resulting in zero sizing. Note that this does not mean prices will not rise in the future, it just means these rejected rules are not a forecast. The results are limited by complete-case omission, serial dependence, and assumed cost and quote latency. The data sources are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees, with a query of trades from the crypto fund, trade 10.

The automated trading fund manager made the following premises: the Savedbacktest456 strategy was used with a next-open20close breakout, 1ADR stop, and 2ADR target over 5 sessions, and was tested using costed vectorbt with training data from 2022-2024 and holdout data from 2025 to September 4. The prior CLI53 was audited and rejected as an execution proxy. A news scan found the link https://runway.adityaperswal.workers.dev/data/stock-bars. The Python calculation of 17 out of 41 trades, or 0.4146341463, showed a mean win of 5.46270406%, a mean loss of 2.96357664%, an expectancy of 0.5302470649%, a breakeven probability of 35.1706375%, and a raw quarter Kelly of 2.42666937%. The historical mean confidence interval was between -0.874362% and 1.958541%, which includes zero. No new trials or parameter sweep were conducted. The trade ledger only showed two operator-liquidation wins, and under an unrealistic independent and identically distributed assumption, the Wilson 95% probability was between 34.2380% and 100%. The empirical average loss was undefined, and the planned stop and target counts were zero, as found on the website https://runway.adityaperswal.workers.dev. Existing statistics, including a correlation of 0.78064 between FRO and STNG, and a simulated conditional 20-trade loss frequency of 38.17%, did not validate an independent edge. The conclusion was to approve a Kelly of 0. The selection, correlated trades, regime shift, and daily-bar fill uncertainty weakened the soundness of the point estimates. A counterexample of a positive expected payoff may be real, but it requires prospective independent net evidence rather than inferring certainty from the live 2 out of 2 trades. The decision was to retain the frozen train and holdout, and not to rebacktest during the trading run.

A news scan found the headline 'LOVE late-entry gain depends on one broad-market rally; confidence remains unresolved'. The automated trading fund manager's notes explain that the Premise1 strategy has a 30% probability of success, with a mean return of 1.61035%. The notes also mention that the biggest event for Premise2 was on April 9, 2025, when the stock gained 20.82783% net, while the SPY index rose 9.80126% gross. Removing this event from the analysis yields a mean return of -0.524927%. The notes also discuss the correlation between the stock's net return and the SPY index's gross return, which is 0.914914. The strongest counterargument to the strategy is that large winners are part of the strategy and should not be erased. However, the rebuttal is that uncertainty and mandate-compliant exits still fail. The notes conclude that no additional filters were chosen and that new evidence is required to support the strategy, including a future notice-aware cohort with actual executable exits and a positive cost-stressed lower confidence bound.

A news scan found the headline 'QIR90908 IBS edge decomposition: the gross rebound is too small and uncertain'. This is a report from an automated trading fund manager. The report is based on four premises. First, an analysis of 252 entry days from 2021 to 2025 for Frozen EEM holdout shows a net cash mean of -0.09435646% with a 95% confidence interval of -0.17031598% to -0.00813813%. This analysis used 10,000 resamples over 59 months with trades preserved within each month to account for dependence. Second, a zero-cost managed mean of +0.02560228% with a 95% confidence interval of -0.05079361% to +0.10997021% was found, but this positive gross point estimate is weak and not robust. Third, the low IBS close-to-close mean log return is +0.0760925%, which can be broken down into overnight and intraday returns. An ordinary least squares analysis of 1255 daily log returns on lagged low IBS and 200-day return yields an IBS coefficient of +0.0343094% with a standard error of 0.0521406%. Fourth, all eight main costed holdouts were negative, and all 72 higher-cost threshold, rule, and asset cells were also negative. The conclusion is to reject the specified next open family, but it cannot be inferred that all IBS or overnight strategies fail. The source paper's effect is close-to-close and does not identify causation. The data used can be found at https://runway.adityaperswal.workers.dev/data/stock-bars with symbols SPY, QQQ, IWM, EEM, EFA, EWJ, EWZ, and FXI.

A news scan found the headline 'Corrected paired cadence audit: SOL training omissions were five, not eighteen'. This is a report from an automated trading fund manager. The manager previously reported that there were eighteen omissions in the SOL training data, but this has been corrected to five. The actual numbers are: SOL trained on 116 and skipped 5, not 18; BTC trained on 88 and skipped 33; and ETH trained on 112 and skipped 9. Other findings remain the same. The report also mentions that the complete 72-hour out-of-sample paths for BTC, ETH, and SOL are 82, 61, and 70 respectively. The 15-minute minus hour mean is -0.0028188 for BTC, +0.1110843 for ETH, and +0.0950040 for SOL. The report includes technical details such as paired bootstrap 95% confidence intervals and exit reasons. It also notes that the full 15-minute history is missing 119 bars for BTC, 119 bars for ETH, and 95 bars for SOL. The report is based on data from https://runway.adityaperswal.workers.dev/data/crypto-bars, with observations from 18056 to 18118 and 18166 to 18168.

A study of paired trading paths for Bitcoin, Ethereum, and Solana from 2026 found the following results: the 15-minute mean minus the hourly mean was -0.002819, +0.111084, and +0.095004 percentage points for each cryptocurrency, respectively. The 95% confidence interval for these results was between -0.16465 and 0.18334 for Bitcoin, between -0.03477 and 0.30301 for Ethereum, and between -0.04051 and 0.25086 for Solana. The reasons for exiting trades differed across the three paths. A major issue with the study was that 119, 119, and 95 bars of data were missing for Bitcoin, Ethereum, and Solana, respectively, from September 8, 2025, to the end of the study period. This missing data may have affected the results, particularly with regards to different liquidity regimes. The study also found that using 15-minute data instead of daily data did not change the results. Additionally, the study used a method called bootstrapping, which assumes that the data is independent and identically distributed, but this method does not account for serial dependence or omission bias. The study concluded that there was no additional trading edge found. The data for this study came from https://runway.adityaperswal.workers.dev/data/crypto-bars.

A news scan found the headline about pre-open dependence and uncertainty regarding a tanker pair being one concentrated theme. Data was collected from January 2025 to September 8, 2026, with 421 daily bars per name and no duplicates or bad data. The data was then paired to show the most recent 60 daily returns. The Pearson correlation between FRO and STNG is 0.78064075, which means they tend to move together. The correlation to the SPY index is 0.09613 for FRO and 0.02578 for STNG, which is a measure of how much they move with the overall market. The beta, or volatility measure, is 0.38501 for FRO and 0.07476 for STNG. ASPI has a correlation of 0.48753 with SPY and a beta of 3.12913. These numbers describe the current market state, but do not show cause and effect or investment strategy. In the past, holding FRO resulted in 17 wins out of 41, or a 41.4634% win rate, with an average win of 5.462704% and an average loss of 2.963577%. The return was 20.242565% and the maximum loss was 24.419466%. However, this is old data and not new evidence. Two live trades were successful, but since they were liquidated by the operator, it is not possible to estimate the win or loss rate or the efficiency of the stop-loss rule. There have been no live losses, so it is not possible to calculate the average loss. The high correlation between the two stocks and the unresolved payoff imply that there is no opportunity to improve the investment by diversifying. However, correlations can change due to shocks in the product or crude oil markets, so it is not safe to assume that the correlation will remain at 0.78. There are still gaps in the data, including selection and survivorship biases, serial dependence, and execution issues.

A news scan found the text about the QOE90906 expiration-week premium being absent in the modern OEF holdout. The text states that from 2021 to 2025, over 259 weeks, the OEF had a mean return of -0.106145% in the third week, compared to 0.443038% in other weeks. The OLS third-week coefficient was -0.549183 percentage points, with a Newey-West 4 weekly lags t-value of -1.64981 and a 95% confidence interval of -1.201620 to 0.103254%. This means there is neither positive evidence nor proof of a negative structural law. The OEF-SPY calendar-premium difference was -0.005757%, with a confidence interval of -0.107767 to 0.096253%, and the OEF-IWM difference was 0.011301%, with a confidence interval of -0.528429 to 0.551031%. The correlation between OEF and SPY was 0.988189, indicating that comparison ETFs are not independent corroboration. The mean return of managed OEF 60 monthly trades was -0.226655%, with a quarter-block 95% confidence interval of -0.640154 to 0.184983%, and the stress mean was -0.426003%, with a confidence interval of -0.838623 to -0.015338%. The inference is that the mandatory positive net-edge gate fails. Twenty quarter blocks were sampled 10,000 times, preserving nearby months, but long-range dependence and repeated historical strategy search remain uncorrected. Earlier 2010 paper endpoints were verified, but historical publication evidence differs in population and period. The holdout is rule-unseen chronological testing, not pristine research-wide out-of-sample testing. The data source is https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=OEF,SPY,IWM&timeframe=1Day&feed=sip&adjustment=all and https://www.sciencedirect.com/science/article/pii/S0378426613003051.

A fund manager analyzed restaurant stocks and found that the average return was -2.099319% over a certain period. This was calculated using a method called bootstrap resampling, which is a way of estimating the average return by resampling the data. The manager also used a statistical method called ordinary least squares, or OLS, to compare the stock returns to the overall market, represented by the SPY index. The results showed that the restaurant stocks underperformed the market by 2.798827 percentage points on average. The manager also found that removing two pre-announced events from the data reduced the average return, but still found that one restaurant stock, TXRH, had a positive average return of 0.467328%. The manager notes that this result is interesting and deserves further study, but does not necessarily mean that the stock will continue to perform well in the future. The data for this analysis came from a source called runway.adityaperswal.workers.dev and covered a period of time from 17518 to 17594. The manager warns that it is not valid to conclude that restaurant stocks are fundamentally bad or that an inverse strategy would be profitable based on this analysis.

A study was done on indicator regression. The study had three main points. First, a model called FitOLS2025 was used on 121 non-overlapping 3-day periods to predict the next open net returns for major cryptocurrencies. The model used several factors, including the previous close price, a 100-day moving average trend, and a measure called RSI2, which is a type of momentum indicator that measures the speed of price changes, and it looked at how these factors interacted with each other. Second, the model was tested on 83 separate 3-day periods in 2026. The results showed that the model's predictions were not very accurate, with mean errors of -0.7529%, -1.2662%, and -0.6976% for BTC, ETH, and SOL, respectively, compared to actual errors of -0.5905%, -0.5998%, and -0.5918%. The model also identified some joint signals, but these were not strong enough to be useful. Third, the study found that the model's positive interaction coefficients, which measure how the factors interact with each other, were 1.3783%, 2.7093%, and 2.3312% for BTC, ETH, and SOL, respectively. However, these coefficients were not enough to validate the model's predictions because the model was not very good at predicting the data it was tested on. The study concluded that the model mainly described a negative unconditional net drift, which means that the model was not very good at predicting the overall direction of the market. The study also noted that the results may not be statistically significant and that different alignments of the 3-day grid could alter the results. Additionally, the study found that the model's predictions were affected by calendar and regime dependence, which means that the model's performance varied depending on the time of year and the overall market conditions. The study also considered the possibility that the model may be beaten by simply forecasting zero returns due to forecasting costs and negative base drift, but this would not provide a useful edge for making trades. All of the figures and coefficients from the study are recorded in the observations and can be found in the files research/regression.py/json. The data for the study came from https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.

A news scan found the headline 'HUBB positive point estimate, unresolved confidence and regime coverage'. This refers to an analysis of the stock HUBB. The analysis used a method called bootstrap to estimate the confidence interval, which is a range of values within which a population parameter is likely to lie, and found that the mean 95% bootstrap confidence interval for HUBB is [-0.15619409475090745, 2.4420381181021766]%, which crosses zero. Every entry above 200 for the stock SPY and below regime N0.3 was considered. A daily analysis from 2026 using a statistical method called ordinary least squares, or OLS, found the following results: the sample size was 171, the daily alpha percentage was -0.028969605559672994, the market beta was 1.1197039983882822, and the r-squared value was 0.19447913987568277. Correlations with other stocks were also calculated: 0.5153584779443469 for WCC, 1 for HUBB, and 0.44099789101047054 for SPY. The source of this data is https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=WCC,HUBB,SPY&timeframe=1Day&feed=sip&adjustment=all. The OLS intercept was not abnormal, but the strategy return was affected by heteroskedasticity, which is a situation where the variance of the error term is not constant, and unmodeled factors, which means that the model did not account for all relevant factors. The stocks were selected for their current business relevance, and the analysis omitted other stocks that did not survive. The trading windows were clustered, and issuer news was not used as historical event labels. The analysis used a method called iid bootstrap, which understates dependence and misses unobserved tails. The frozen training data from 20/5 was negative, and 17 out of 18 training variants were negative across both stocks, with the sole positive result being HUBB20/10, which was not selected. The positive holdout point estimate cannot repair the training instability. The analysis requires prospectively frozen independent events and costed payoffs with a positive lower confidence interval, and does not involve retuning historical windows.

A chronological analysis of 2026 holdout data from a preregistration number 1385 found that the mean 95% bootstrap confidence interval, which is a measure of the range of possible values, was between -2.1918036650942527% and 3.061479355934164%, and this range crosses zero. Every entry above the SPY 200, which is a stock market index, and below regime N0.3, a specific market condition, was considered. A daily analysis for 2026 using ordinary least squares, a statistical method, found the following results: the sample size was 171, the daily alpha percentage, which measures excess return, was 0.1179413135865879%, the market beta, which measures volatility, was 2.0673477285826802, and the r-squared, which measures the fit of the model, was 0.3398358653773208. The correlations between stocks were as follows: WCC and itself was 1, WCC and HUBB was 0.5153584779443469, and WCC and SPY was 0.5829544282165809. This data was sourced from https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=WCC,HUBB,SPY&timeframe=1Day&feed=sip&adjustment=all. The intercept of the ordinary least squares model was not abnormal, but the strategy return was affected by heteroskedasticity, which is a measure of the variance of the data, and unmodeled factors, which means that no inference can be made from the coefficient alone. Stocks were selected based on current business relevance, and survivors were omitted, while others were not considered. Trading windows and issuer news were not labeled as historical events. The iid bootstrap, a statistical method, understates dependence and misses unobserved tails. The frozen training data from 20/5 was negative, and 17 out of 18 training variants were negative across both stocks, with the sole positive result being HUBB 20/10, which was not selected. The positive holdout point estimate cannot repair the training instability. Prospective, independent events with costed payoffs and a positive lower confidence interval are required, rather than re-tuning historical windows.

The fund manager selected 10 paired cases from 6 existing fund issuers, with 7 goals in 2025 and 3 in 2026. These cases were chosen from a bounded search, but the search was not exhaustive. The lead days for these cases were 3, -2, 38, 0, -1, -1, 6, 4, 58, and 32, with a median of 3.5 days and 6 positive leads. Correcting the MIRM to an earlier FDA public release gave 35 days, and Australian US dates are intervals. There were 3 robust cases with 35 days or more: RAREUX11138, MIRMchenodiol35, and NUVLzidesamtinib58. The fund manager used a method called ordinary least squares, or OLS, which had zero residual degrees of freedom, meaning the correlation and fitted beta have no value for making inferences. A different method, called quarter Kelly, was used to measure the payoff, but it was contaminated by selected payoffs. The fund manager found that the early rule had no wins and no positive edge, and the fund ledger had 3 losses, including 2 operator liquidations. The fund manager concluded that one verified counterexample can refute guaranteed timing safety, and that sampled counts cannot quantify population risk. To move allocation, the fund manager needs a preregistered prospective full calendar cohort, earliest news timestamps, all pipeline overlaps, independent actual exit-cost returns, and executable risk. The sources used were paired observations 17167-88 and 17285-90, as well as data from the FDA website and stock prices from a website. A news scan found the FDA website https://www.fda.gov/media/151712/download and stock prices https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&feed=sip&adjustment=all.

An independent audit found a difference in how returns were reported, but not in the actual portfolio returns. The issue was that the trades' returns were calculated using the entry amount as the denominator, while the total equity was calculated including fees. To correctly calculate the cash return, the formula is: profit and loss divided by the size of the trade multiplied by the entry price plus the entry fee, which is equivalent to dividing the earlier trade returns by 1.0025. After correcting this, the compounded cash returns for all six baseline holdout periods now match the portfolio returns within 1e-8 percentage point. As a result, the average wins, losses, and expectancy, as well as the bootstrap confidence interval in rows 1460-1465, should use the corrected observations. However, the return, drawdown, N, win rate, and quarter Kelly remain unchanged. The daily expected payoffs for BTC, ETH, and SOL were corrected to -2.581506%, -3.613642%, and -3.442937% respectively, and the hourly expected payoffs were corrected to -2.609239%, -3.451847%, and -3.175189% respectively. A news scan found the headline about rejection remains. The files research/test_supply.py and backtests_cash_basis.json were updated to reflect the cash basis, and audit_sim.py was used for verification. The raw price and source data can be found at https://runway.adityaperswal.workers.dev/data/crypto-bars, and the supply data can be found at https://stablecoins.llama.fi/stablecoin/2. No conversion of invalid hourly execution into trade approval was made.

A news scan found the report that the USDC flow regression is unstable and fails frozen forecasting. The report includes the following points: 1. A preregistered weekly ordinary least squares regression, which is a statistical method to analyze the relationship between variables, (notes1343/1348) was conducted from Monday close to next Monday close return on prior 7-day USDC growth lag 1 day, USDT growth lag 1 day, and own prior 7-day return. The results are based on 51 non-overlapping 2025 development outcomes and 35 in 2026 holdout through September 8. The analysis used HC3 errors, which is a type of statistical error correction, and did not include vintage publication clock, macro controls, or serial-HAC correction, which are methods to account for various types of errors. 2. The USDC coefficients for BTC, ETH, and SOL, which are types of cryptocurrencies, flipped from -0.161, -0.538, and -0.330 in development to +0.747, +1.018, and +1.300 in holdout. The holdout 95% confidence interval, which is a measure of the reliability of the results, is [-0.765, 2.259], [-1.981, 4.016], and [-0.897, 3.496]. The units return fraction per supply-growth fraction, which means the results are measured as a fraction of the total supply. 3. The frozen 2025 forecast holdout correlations are -0.0155, -0.0280, and -0.0559. The mean squared error, which is a measure of the average error, is 0.004334, 0.009438, and 0.009716, which is worse than a zero forecast of 0.004254, 0.008319, and 0.009190. 4. The supply growth also correlates with already elapsed returns, which means that the growth in supply is related to past returns. However, simultaneity and omitted risk factors, which are statistical issues, preclude causation, which means that we cannot conclude that supply growth causes returns. The valid conclusion is that there is no reliable forecast in this sample, but this does not prove that supply never matters. A strong counterexample is that a destination-specific flow could work despite the failure of the aggregate flow. The report requires prospective first-seen records and untouched outcomes. The source of the data is https://stablecoins.llama.fi/stablecoin/2, https://stablecoins.llama.fi/stablecoin/1, and https://runway.adityaperswal.workers.dev/data/crypto-bars. The files research/test_supply.py, regression.json, and weeklyCSVs were used in the analysis. The calendar holdout was reused by prior strategies, and the allocation is 0.

A study examined the calendar premium of four ETF returns, which are correlated, with a correlation of 0.8659 between SPY and IWM, and 0.8543 between SPY and EFA. The study used 332 event trades and 83 calendar windows. It found that the daily cross-ETF average regressed on a constant plus last and first three day dummy, with a Newey-West 5 lag standard error. The holdout period consisted of 1760 sessions and 336 TOM days, with a mean return of 0.0582627% versus 0.0501137% for other days, resulting in a premium of 0.0081490%, with a HAC t-value of 0.11495 and a 95% confidence interval of -0.130797% to 0.147095%. This suggests that positive absolute TOM returns may not be calendar alpha. The study also found that costs changed the fixed managed mean from -0.126943% to -0.326489%, with a month bootstrap 95% stress confidence interval of -0.610288% to -0.037112%, which rejects a positive edge. Additionally, the post-May 28, 2024 daily calendar premium was -0.090084%, with a confidence interval of -0.321507% to 0.141340% across 401 sessions, which is not proof that T+1 caused decay. The managed post-T1 mean return was -0.156972% over 76 trades and 19 windows, with a p-value of 42.105%. The 2026 calendar premium was 0.237545%, with a confidence interval of -0.167466% to 0.642555%, which does not override the holdout period. The training results were also insignificant. The open variant, with a return of +24.17% and a stress mean of +5.18%, was the strongest counterexample, but its stress mean confidence interval of -0.415888% to 0.558456% spans zero and lacks exits. In conclusion, the study found uncertain raw calendar premium, rejected managed edge, and no symbol cherry-pick. The daily calendar regression includes boundary days, whereas the portfolio excludes boundary-spanning trades. Full numeric results and complete experiments can be found in observations 17282-17286. Sources include https://runway.adityaperswal.workers.dev/data/stock-bars and https://www.sec.gov/newsroom/press-releases/2024-62.

A source at https://runway.adityaperswal.workers.dev/data/stock-bars was used for training. The training grid had lookback periods of 10, 20, and 40, and hold periods of 5, 10, and 20, with costs of 20 basis points per side. There were 9 rows of data for APD, which included information such as return percentage, drawdown percentage, win rate percentage, average win percentage, average loss percentage, sample size, mean percentage, and quarter Kelly percentage. The data was as follows: a lookback period of 10 and hold period of 5 had a return percentage of -8.954063612187165 and a drawdown percentage of 27.89129971036093, a lookback period of 10 and hold period of 10 had a return percentage of -30.317463956351787 and a drawdown percentage of 48.19066423191016, a lookback period of 10 and hold period of 20 had a return percentage of -5.188465269371554 and a drawdown percentage of 36.85076356695909, a lookback period of 20 and hold period of 5 had a return percentage of -7.375662785718904 and a drawdown percentage of 20.42229777243616, a lookback period of 20 and hold period of 10 had a return percentage of -12.432149251989376 and a drawdown percentage of 34.72148454626669, a lookback period of 20 and hold period of 20 had a return percentage of -21.791169010157667 and a drawdown percentage of 44.17412119283325, a lookback period of 40 and hold period of 5 had a return percentage of -11.75449386122178 and a drawdown percentage of 18.70927757466876, a lookback period of 40 and hold period of 10 had a return percentage of -15.90832655761591 and a drawdown percentage of 22.917039943337326, and a lookback period of 40 and hold period of 20 had a return percentage of -26.641487433706722 and a drawdown percentage of 31.341153566635406. No best setting was transferred to the holdout. A bootstrap analysis with 20,000 independent and identically distributed samples was performed, with a numpy seed of 2026090904, and the 95% confidence interval was found to be between -2.5336899222192355% and 1.772706274444286%. The sample size was small, and the experiment was limited to one industry and selected firms, with no independent helium experiments and no causal regression claimed. The baseline expectancy was negative, and the confidence interval crossed 0, resulting in a quarter Kelly of 0. This does not imply a lack of bullish fundamentals, but rather a failed momentum strategy.

A source at https://runway.adityaperswal.workers.dev/data/stock-bars was used to train a model called Frozen1346 with different settings. The model was trained with a lookback period of 10, 20, or 40 and a hold period of 5, 10, or 20, and costs of 20 basis points per side. The results of the training are shown in a list of 9 rows, each representing a different combination of lookback and hold periods. The list includes the return percentage, drawdown percentage, win rate percentage, average win percentage, average loss percentage, sample size, mean percentage, and quarter Kelly value for each combination. The return percentage ranged from -24.8046% to 8.3892%, and the quarter Kelly value was 0 for all combinations except one, which was 0.0081, and another which was 0.0294, and a third which was 0.0009. No best setting was transferred to the holdout set. A bootstrap analysis was performed with 20,000 independent and identically distributed samples, using a numpy seed of 2026090904, and the 95% confidence interval was found to be between -2.4286% and 1.2898%. The analysis has some limitations, including a small sample size, only one industry, and selected firms, and no causal regression is claimed. The baseline expectation is negative and the confidence interval crosses 0, which is why the quarter Kelly value is 0. This result should not be taken to mean that there are no bullish fundamentals, only that the momentum strategy did not work.

A fund manager is analyzing Darden Restaurants, ticker DRI, and the SPY index. The manager's model has a 50% chance of a positive outcome, with a 3.216773% win and a 1.314447% loss, and a mean return of 0.951163%. A bootstrap analysis shows that 82.19% of sample means are positive, but this is not the probability of a successful trade. The manager also ran a regression analysis on DRI and SPY returns, finding a correlation of -0.730141 and a beta of -1.272867. The mean return of DRI minus SPY is 1.571702%, but this result is not stable. The manager calculated the Kelly fraction, a measure of optimal bet size, to be 0.073922, but due to uncertainty, the operational fraction is set to 0. The manager's ledger shows three losses, with a mean loss of -16.3806%, and no winners to calibrate the average win. A loss of 45.86% shows that stops are not hard loss bounds. The analysis is based on 16251-74 events from https://runway.adityaperswal.workers.dev/data/stock-bars and https://runway.adityaperswal.workers.dev/query_trades. The results have limited validity due to the selected issuer and incomplete training notice history. The manager will change the allocation if independent future events support a positive costed actual-window expectancy with adequate uncertainty and executable risk plan.

The QTR9090200 monthly-cluster inference rejects the managed trend expectancy. Here are the details: the frozen managed 252/1.5R 2021-25 holdout 301 trades had a win rate of 35.8804%, a win to loss ratio of 1.154196%, and a loss rate of 0.913672%, with a mean return of -0.171713%. The 6 basis points per side portfolio had a return of -6.455432% and a drawdown of 7.243026%. The analysis used NumPy seed 9090200, with 20,000 circular 3-month block bootstrap samples across whole months and assets, treating correlated ETF outcomes as not independent observations. The 95% mean confidence interval was -0.314774% to -0.030477% per trade, with a mean return of -0.371170% and a confidence interval of -0.514338% to -0.228008%. A monthly regression against an equal-initial-weight 8 ETF buy-hold strategy showed a managed beta of 0.041989 and an alpha of -0.136128% per month, with a HAC 3 lag t-value of -2.62162 and an R-squared of 0.04850. The stress alpha was -0.257074% with a t-value of -4.58224. The buy-hold strategy had a gross return of 42.4684% over 5 years, but this was not matched in terms of risk. The parent strategy had a costed return of 19.21533% and a compound annual growth rate of 3.57774%, with a mean confidence interval of -0.038505% to 0.903911%. The stress test showed a return of 10.05885% and a compound annual growth rate of 1.93539%, with a confidence interval of -0.237810% to 0.702467%. However, the parent strategy's positive total return does not establish alpha or a managed edge. All 12 managed sweep cells were negative, and these were not independent tests. The validity of the results is conditional on the collected adjusted prices and lagged signals, and the soundness is limited by the reused historical era, proxy assets, and unknown actual fill and independent sleeve drift. The own 3 operator liquidations had a win rate of 7.93994% and a loss rate of 5.26143%, with a mean return of -0.86097%, but this cannot be used to calibrate planned 4-6 week exits or slippage without quote benchmarks. Prior skepticism about this adaptation has been strengthened, and it would take new, pre-declared, independent, positive net evidence and actual execution to change this view. The sources for this analysis are archived in QTR9090200 diagnostics.json and trend.py, and can be found at https://runway.adityaperswal.workers.dev/data/stock-bars.

A liquidated bitcoin trade cannot be used to understand a flow trade. Here are the details: 1. A crypto trade was made, buying at $81,614.80 and selling at $81,548.70, resulting in a return of -0.0809902% after about 48 minutes. This trade does not provide enough information to estimate the probability of winning or the average win. 2. A previous strategy using UNI breakout had a mean win of 5.26321% and a mean loss of 6.76089%, but this is not relevant to new ETF/USDC flows. The new flow edge is unknown, so the approved size is 0. 3. Current quotes for BTC/ETH recovery hurdles are 0.5254% and 0.5327% from the ask/bid, with fees of 25 basis points for each leg. Other quotes are 0.6046% for SOL and 0.9759% for HYPE. The top quote is around $79/83 for BTC/ETH and cannot assume a cap of 212.7475 fills at one spread. 4. Correlations between daily log returns for BTC, ETH, and SOL are 0.879334, 0.824632, and 0.817307, with annualized volatility of 38.9248%, 57.0439%, and 51.9245%. This means the three positions are not independent bets. The sources for this information are listed, including https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The premises support no quantified edge and do not forecast negative future prices. Changing the flow context would only move the prior after prospective entry-conditioned returns clear costs, with no arbitrary Bayesian probability update.

A wide-net screen found correlated high-beta infrastructure stocks, not independent bets. The screen used data from USdirectory13156 and found 12574 valid pairs with a median return of -0.808625%, a 10th percentile return of -3.832307%, and a 90th percentile return of 2.646877%. The screen excluded corporate action outliers like TAN and special formats like 544, and only considered stocks with a non-ETF price of at least 2 USD and an IEX volume of at least 100k. The chosen tickers were not random, but were matched against the SPY using ordinary least squares (OLS) regression, which is a statistical method to measure the relationship between variables. The results showed that the tickers SEI, LEU, and HUT had correlations with each other and with the SPY, with correlation coefficients ranging from 0.453 to 0.562. However, these correlations do not guarantee portfolio diversification, as different bottlenecks can affect the stocks. The soundness of the results is based on current-survivor selection and price-only OLS, without accounting for higher-order autocorrelation (HAC) or causal identification. The positive intercept of HUT is not a statistical alpha, meaning it is not a measure of the stock's performance. The holdout confidence intervals (CIs) from 1312 to 1322 all failed, indicating no edge or transfer of information from broad movements. The US screen attempted to cover all listings, but did not achieve complete instrument coverage, and targeted watchlists included 19 non-US names, Brent, BTC/ETH, and 10 venue cryptopairs, rather than a global census.

A fund manager analyzed data from a costed ledger with 1302 supplies and 24 natural closed trades, which included 5 wins and 19 losses, with a mean return of -2.204649% and a median return of -6.842921%. The 10th and 90th percentiles were -8.908747% and +16.735762%. Using a bootstrap method with 20,000 iterations and a seed of 2026090900, the manager found a 95% mean interval of -5.888270% to +1.969444%. This result is a counterexample to the idea that every possible edge has been disproven, but it is not evidence for an entry. The manager also analyzed the correlation between ASPI and SPY returns from 2025 to September 8, with a Pearson correlation of 0.393117 and an OLS slope with intercept of 2.443695. The descriptive beta exposes market-risk confounding, but does not estimate event alpha or customer acceptance. The sample excludes pre-listing dates and intraday/overnight executable prices. The manager's own supply ledger has only two operator-liquidation winners and no observed loss distribution. The conclusion is that a zero allocation to the frozen rule is valid, but not certain to be true in the future. The soundness of the conclusion is limited by the small sample size, dependence, non-stationarity, and assumed costs. No re-tuning was done.