The Chewy event timing audit is based on three premises: a forecast needs the return window that actually contains the announcement, event-gap history is a distribution estimate and not an option return backtest, and a two-event holdout cannot establish tail calibration. There were 18 individual issuer notices from March 2022 to May 2026, with nine after-close releases through March 20, 2024, and nine before-open releases thereafter. The audit mapped after-close announcements to the next-session open versus the announcement-day close, and before-open announcements to the same-session open versus the prior close. All 18 exact dates, original notice URLs, publication dates, aligned prices, and log returns were saved in observations 7151-7168. The primary notice established the intended phase, and historical first-news ticks and any early preannouncements were not independently reconstructed. The audit used adjusted IEX daily bars from https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=CHWY&timeframe=1Day&start=2022-01-01&end=2026-09-05&adjustment=all&feed=iex, which included 1172 bars. Sixteen events from 2022 to 2025 were used for training, while March 25 and June 10, 2026, were held out before calculation, with no 2026 refit. The calculation used the formula J = ln(event open/preclose). The training results showed an RMS of 10.3956%, a mean absolute of 9.2054%, and a median absolute of 9.0425%, with a close-to-close RMS of 13.0700%. An IID event bootstrap with 10,000 seed 2026090608 showed an RMS 95% interval of [8.0483, 12.4958]%. The held-out gaps were +10.1067% and -0.0490% log, with an RMS of 7.1466% and a mean absolute of 5.0779%. The current before-open phase was verified by the issuer on August 19, https://investor.chewy.com/news-and-events/news/news-details/2026/Chewy-Announces-Fiscal-Second-Quarter-Financial-Results-Conference-Call/default.aspx. The audit found that the deliberately wrong all-before-open mapping gave an RMS of 6.1948%, understating variance-window exposure. The results were insufficient to validate event tails or distinguish competing forecasts, with no p-value or trading edge claimed. The evidence that would move the decision includes prospective independent event/option bid-ask returns with uncertainty excluding zero after costs, plus execution entitlement. A counterexample to transportability is changing margins, financing, or preannouncements that change surprise variance, and opening IEX prices can differ from consolidated auction prices. The timing logic is valid, but empirical adequacy for a paid tail bet remains unproven.
A news scan found the Sea-Intelligence summary from September 2, which covers July 2026. The reliability was 56.4%, down 6.1 percentage points, and the mean delay was 6.06 days, up 0.59 days, but only for vessels that arrived late. This is called a conditional delay. To understand this, we need to know that conditional means based on a condition, in this case, late arrival. From the same release, we can derive that in June, the reliability was 62.5% and the mean delay was 5.47 days. Using these numbers, we can calculate a diagnostic, which is a measure to help understand the situation, and it increased by 28.81%. This index does not show the full picture of lateness because it does not account for vessels that arrived on time or early, and it does not consider the weight of the vessels. Maersk cited this statistic in a weather notice on September 4, but it is not independent confirmation because it comes from the same origin. Hapag-Lloyd reported different delay ceilings for selected ports on September 3, but this is not a contradiction because it is a different population and period. The collector index labels June as 62.6%, which is 0.1 percentage points different from the inferred 62.5%, and this difference is preserved because it could be due to rounding or revision. The raw data, including AIS, counts, paired arrivals, and spreadsheets, is not available, so no regression, confidence interval, or causal storm coefficient can be calculated from these aggregates. Previously, it was thought that physical disruptions were likely to have a significant impact, and this is corroborated by current carrier restrictions, which strengthen the mechanism, but do not affect the probability of winning stocks. New accepted invoices and independent throughput would strengthen the cash hypothesis, while normalizing delivery metrics would weaken it. The sources for this information are Sea-Intelligence and Maersk and Hapag-Lloyd websites.
The automated trading fund manager used prior measured range inputs of 3.663% for Bitcoin, 4.840% for Ethereum, and 6.386% for Solana. These are not new trade targets. The manager considered a 25 basis point fee on each side for credited assets, as documented, and assumed an additional 10 basis point slippage on each side for research stress. The formula to calculate the impact of these costs is k = (1 - 0.0025)^2 * (1 - 0.001) / (1 + 0.001) = 0.9930182255. The manager then calculated the net loss and net win for each asset: Bitcoin had a net loss of 4.3356% and a net win of 4.7580%, Ethereum had a net loss of 5.5044% and a net win of 6.5111%, and Solana had a net loss of 7.0396% and a net win of 8.8139%. The risk-reward ratio was 1.0974 for Bitcoin, 1.1829 for Ethereum, and 1.2521 for Solana. To reach a net 1.5 risk-reward target, the manager would need 7.2522% for Bitcoin, 9.0177% for Ethereum, and 11.3367% for Solana, but this is not permitted without new measured payoff validation. The manager also considered a 1% equity risk cap, which would be $196.28 for Bitcoin, $154.60 for Ethereum, and $120.89 for Solana. A news scan found the snapshot of best ask prices to be $79.45 for Bitcoin, $83.55 for Ethereum, and $100.94 for Solana, and bid prices to be $30 for Bitcoin, $20 for Ethereum, and $99.59 for Solana, which underscores that fills are not guaranteed. The fee-on-received-asset convention used here differs slightly from the fee-on-cash-entry ledger, and exact account activities are still required before execution. Lower maker fees could reduce costs but may limit fill probability and increase adverse selection, but this has not been tested. No new strategy backtest or independent edge is claimed.
A review of the Catalyst trades from September 2-4 was done using numpy. The average profit per trade was -16.3806%, with 0 wins out of 3 trades, and a total profit and loss of -31.937258, which matches the venue's record of -31.94. Two trades were stopped by the operator, which affected the intended strategy outcomes, including one RARE stop exit with a loss of 45.8601%. Since there were no winning trades, it is not possible to calculate the average win, and it should not be assumed that future trades will have a 0% chance of winning. As a result, the live trading record cannot be used to calibrate the Kelly criterion. A separate analysis of CHWY events was saved, which showed a net return of -1.719988% and a clipped quarter Kelly of 0. The historical loss stress exceeded the account risk cap, but this does not guarantee that future losses will be limited. The premises of the trading strategy, including the edge and time window, have not been validated, which means that the trade size should be 0. A future rally would not prove that the strategy has an edge. The data sources used include the execution ledger query_trades fund, historical stock bars, and dates from Chewy's financial reports. No new backtest or Monte Carlo simulation is claimed, and resampling the 3 censored trades would not provide any new evidence.
A news scan found the text describing a study on trading strategies. The study used a fixed six-sector panel and adjusted daily stock prices from the IEX exchange from 2020 to 2026. The training period was from 2020 to 2023, with usable strategy trades starting in February 2021, and the holdout period was from 2024 to September 4, 2026. The study found that the six daily return series had a mean pair correlation of 0.570715 during training and 0.398842 during the holdout period. The correlation-eigenvalue participation ratio was 2.15466 and 2.87526, which means the descriptive dimensions were not effective for independent trade counts. The baseline had 647 training trades and 598 holdout trades, which clustered in 132 and 126 occupied entry weeks, respectively. The study used a method called bootstrap resampling, which involves re-sampling the data, to test the robustness of the results. The study found that many trades did not result in a positive payoff, with governing four-week mean intervals of -0.21028% to 0.28597% for the baseline, -0.47654% to -0.02183% for the trend, and -0.57805% to -0.02514% for the trend low volatility. A regression analysis was used to examine the relationship between trade returns and factors such as trend and low volatility. The holdout coefficients were 0.86763, -1.16261, -0.96635, and 1.02668, with t-values of 2.804, -3.358, -2.383, and 2.346, respectively. However, the study found that these results were unstable and should not be used to make trading decisions. A strong counterexample was found, with a downtrend and high volatility resulting in a holdout mean of 0.86763% over 150 trades, with a one-week confidence interval of 0.26937% to 1.47896%. The study concluded that there was no untouched evidence to support selecting this particular trading strategy, and it should be recorded as a discovery rather than used for trading. The study's results were limited by the lack of incident labels and the potential for week clusters to share multiweek shocks. The data source was https://runway.adityaperswal.workers.dev/data/stock-bars.
A review of trades for a quant fund found three closed positions in PEAD, all of which were exited by an operator. The returns from these trades were -5.39236% for SNOW, +7.93994% for DELL, and -5.13050% for GTLB, resulting in a total loss of approximately $1.04605. The win rate was 1 out of 3 trades, with an average winning return of 7.93994% and an average losing return of 5.26143%. These trades lasted for 0.87414, 1.80126, and 1.96827 days, which is shorter than the expected 4-6 weeks. Since there are no reference quotes or benchmarks, it is not possible to estimate the impact of fill slippage. The review concludes that these trades do not provide enough information to calibrate the overnight strategy or determine the odds of natural exits. The data was retrieved from query_trades on September 6, 2026, and is labeled as observation5796.
A study analyzed five exchange-traded funds, or ETFs, with identical 2684-session histories from January 4, 2016, to September 4, 2026. The data, which includes open, high, low, and close prices, or OHLC, was split into two time periods: 2016-2020 and August 2021 to September 4, 2026. The SPY ETF had a gross mean overnight return of 3.373656 basis points, with a geometric breakeven point of 1.587993 basis points per side. The study found that the gross compound return of 56.98655% became -11.04239% at 2 basis points per side and -90.82762% at 10 basis points per side. A statistical analysis, called HAC 5 OLS, of the next day's open and close prices based on current and past intraday and 20-day volatility data showed a lagged beta of -0.015328, t-value of -0.620745, and R-squared of 0.02267. This means that price-only predictors did not identify signed flows. The strongest counterexample was the IWM ETF with a low cost of 15.96% and the SPY ETF with a lagged return of 6.38%, but their confidence intervals crossed zero and the base cost failed. The study also found that the overnight and intraday benchmarks exactly reconciled when corrected for passive close-to-close returns. The study's validity is limited by auction and dividend feed semantics, previously inspected dates, and bootstrap stationarity. The data source is https://runway.adityaperswal.workers.dev/data/stock-bars, with additional information in statistics.json and diagnostics.json.
A news scan found the text describing an analysis of HAIN. The analysis used a method called bootstrap, which is a way to estimate how reliable the results are, with a seed of 60606 and found that the average return was -10.97667%. The results also showed a 95% confidence interval, which is a range of values that the true result is likely to be within, of -23.21708% to 2.91121%. Another calculation, called the Wilson confidence interval, which is a way to estimate the probability of an event, found a 95% confidence interval of 3.00534% to 56.35028%. The analysis noted that the results are not precise because the sample size is small and there may be unknown patterns in the data. The analysis also compared the results to two other stocks, SPY and XLP, and found that HAIN had a lower return. The analysis used a method called ordinary least squares, which is a way to find the best fit line to a set of data, to model the relationship between HAIN and the other two stocks. The results showed that HAIN had a daily return of -0.4030566% and was closely related to SPY and XLP. The analysis noted that the results are based on a small sample size and may not be representative of the entire market. The data used in the analysis came from a source called IEX and covered the period from 2025 to September 4, 2026. The analysis also noted that the results are subject to selection risk, which is the risk that the results are biased because they are based on a selected sample rather than the entire market.
A study tested if a flexible forecast of volatility, a measure of how much prices move, could outperform a simpler baseline forecast. The study used six exchange-traded funds, or ETFs, which are like stocks that track a group of assets, and their daily closing prices from January 3, 2022, to September 4, 2026. The baseline forecast was a combination of three methods: a 21-day average, a 63-day average, and a weighted average that gives more importance to recent prices, known as an exponentially weighted moving average or EWMA. The competitor forecast used a method called HAR, which is like a linear regression, a statistical tool that finds the relationship between variables, to predict future volatility. The study found that the flexible forecast did not consistently outperform the baseline forecast across the six markets. The study used a measure called QLIKE to compare the forecasts and found that the results were not statistically significant, meaning they could be due to chance. The study also found that the flexible forecast did not always perform better than the baseline forecast, even when it made certain predictions that seemed too good to be true. The study concluded that the flexible forecast did not demonstrate a general improvement over the baseline forecast and should not be adopted based on its performance in certain situations. The data used in the study came from a source that provides stock prices, and the study was limited by the noise and shocks in the markets, which made it difficult to make precise predictions.
A review of the Catalyst fund's trades from September 6 found three closed trades with no wins. The returns were a loss of 0.439321% for MIRM, a loss of 2.842420% for IONS, and a loss of 45.860062% for RARE. The average loss was 16.380601% and the total dollar loss was $31.937258 before system costs. Two trades were closed by the operator and one trade had a stop sale. The RARE trade was closed at $13.98976, which is 41.709333% below the stop trigger of $24. This was not necessarily due to order routing issues, as there was no executable quote at the time. All trades were fractional market-notional entries, with no limit comparator and no reference quote to estimate entry slippage. The winner payoff was not observed and the quarter Kelly sizing was 0 by mandate. These three trades should not be grouped together as a sample of earnings events or used to infer a true win probability of 0%. The data comes from the query_trades fund, with a status of closed, and provides information on account 850.99 and fund 49.14, with a 1% risk of $8.5099. This is evidence of execution and outcome, not a causal rule backtest.
A study reconstructed hourly data into daily closes in UTC time, only counting 24-hour days with no fill, and used data from 2024 to 2025 for training, with SOL starting in August 2024. The historical evaluation took place from January 1 to September 5, 2026. There were 484 training data points and 246 evaluation data points for all three premises. The study found that the daily return correlation between BTC and ETH increased from 0.78545 to 0.91234, between BTC and SOL from 0.76891 to 0.88042, and between ETH and SOL from 0.76842 to 0.88022. The evaluation annualized volatility was 46.925% for BTC, 64.509% for ETH, and 64.059% for SOL. The standard deviation on weekends was 1.560%, 2.498%, and 2.685% respectively, which is lower than on weekdays at 2.741%, 3.679%, and 3.595%. However, this does not mean that weekends are safe, as 1% of the tails are still large. The mean daily returns were -0.01126% for BTC, -0.02482% for ETH, and -0.02767% for SOL. The conclusion is that three crypto positions can share one risk driver, and it is not recommended to pool three correlated histories as independent support for the Kelly criterion. One counterargument is that the common-date training covers a shorter SOL history and periods differ, so the correlation change is descriptive, not causal or guaranteed to be persistent. The study excluded missing or incomplete days and used venue-specific feed coverage and prior inspected holdout bound inference. All train and evaluation quantiles, counts, means, standard deviations, and correlations were recorded observations, sourced from https://runway.adityaperswal.workers.dev/data/crypto-bars.
A preregistered note, number 295, analyzed daily training returns from 2023 to 2024, and 250 held-out returns from 2025. The analysis used a 2-factor ordinary least squares method on log returns against the SPY and XLB benchmarks, considering known shock dates from 2026. The holdout predictive R2 was 0.5952 for the linear model and 0.5146 for the APD model, with a root mean square error of 0.7677% and 1.2194% daily. The five-session residual log percentage on March 2, March 19, and July 10 was +0.5278, -1.3739, and -2.0582 for the linear model, and +6.3321, -0.3003, and -1.1084 for the APD model. This analysis is not an executable event backtest because it includes announcement-day close-to-close, related incidents, unknown first-publication minute, and earnings confounds in a 20-day July window. A post-hoc robustness test using the SPY benchmark only changed the linear model results to -3.4158, +1.0627, and -0.8356, and the APD model results to +0.1830, +3.4991, and +0.7979, indicating that attributions depend strongly on the benchmark. State Street's September 3 XLB weights, which were 12.97% for the linear model and 4.76% for the APD model, confirm that the benchmark contains both and cannot be an unaffected control. The weights were not backcast. The conclusion is that no stable causal or equity payoff was identified, and positive physical exposure does not establish a trade edge. The 2025 overlapping placebo windows are a descriptive distribution and not independent nor p-values. No win probability is inferred. The sources of the data are https://runway.adityaperswal.workers.dev/data/stock-bars and https://www.ssga.com/us/en/individual/etfs/state-street-materials-select-sector-spdr-etf-xlb. Every event statistic observation is from 5419 to 5436, and the scripts and results are archived.
The performance of a merger exchange-traded fund, or ETF, is uncertain. The fund's returns were analyzed from January 2016 to December 2021 and then from January 2022 to August 2026. During the later period, the fund had a return of 20.68199%, compared to 15.04870% for another fund called MNA, 19.32620% for a fund called BIL, and 70.66035% for a fund called SPY. The merger fund's average annual return was 4.13930%, which is not much higher than the 3.88576% return of the BIL fund. To estimate the uncertainty of the fund's returns, 10,000 random samples were taken from 21-session blocks of data, which is a statistical method called bootstrapping. The results showed that the difference in returns between the merger fund and the BIL fund could be anywhere from -2.96955% to 3.59355% per year. A statistical model called ordinary least squares, or OLS, was used to analyze the relationship between the merger fund's returns and the returns of the SPY fund. The model showed that the merger fund's returns were not significantly related to the SPY fund's returns during the later period. The fund's performance was also analyzed in terms of its beta, which is a measure of its volatility, or risk, relative to the overall market. The results showed that the fund's beta was low, but this does not necessarily mean that it is not exposed to the risk of a market crash. The analysis also showed that the fund's returns were not significantly different from zero during the later period, which means that the evidence does not support the use of short-term trading strategies for this fund. The data used for the analysis came from a website called runway.adityaperswal.workers.dev, and the original motivation for the analysis came from a research paper published by the University of Chicago.
A news scan found the text: Premises: same stock only one completed FDA deadline; arbitrary stock windows measure background drift, not dated edge. Fixed 15 session non-overlapping windows with 16 session step, IEX 2021–Sep 4 2026, fee 5/slip 10bp each side. Train 61 windows mean 1.2841%, median -1.5474%, win 45.9016%, bootstrap 95% mean CI [-6.0859, 11.4398]%. Holdout 25 windows fully entered/exited 2025+, mean 1.6538%, median 0.8703%, win 56%, CI [-3.6993, 7.1608]%; one Dec 18–Jan 13 boundary window excluded explicitly. 20,000 iid bootstrap resamples, seed 60604; non-overlap reduces mechanical dependence but regime autocorrelation remains. Descriptive 420 daily rows 2025+: corr SRRK/XBI 0.5691, SPY 0.3380; two-factor OLS beta XBI 1.2164, SPY 0.09075, daily intercept -0.06767%, R2 0.3243. No robust SE or causal claim; XBI contains biotech exposure and factors correlated. Conclusion: rising price or positive placebo mean does not identify event premium. Need prospective independently classified manufacturing-resubmission cohort with announcements known before entry and full pipeline overlap. No posterior approval probability assigned; N1 cannot calibrate hazard. Source https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=SRRK,XBI,SPY&timeframe=1Day&feed=iex&adjustment=all. All measured rows/metrics archived.
A news scan found the following information about Netflix stock forecast. The forecast is based on several premises, including independently forecasting physical variance, holding out outcomes, and not equating physical variance or better forecast score to risk-neutral option value. The log-adjusted close return is represented by x. The fixed forecast variance is calculated as the average of 252 times the variance of 21 days, 63 days, and the exponentially weighted moving average of x squared with a lambda of 0.94. The training period is from 2023 to 2025, and the winsor clip bounds are from -0.04953405 to 0.06813260, which are fixed before the 2026 comparison. There are 165 held-out next 5 observed session mean-square labels, and the training evaluation would purge 5 sessions. The current realized volatility over 21 days is 0.3851876, over 63 days is 0.3787668, and the exponentially weighted moving average is 0.3769041. The blended variance is 0.3803027, and the winsorized variance is 0.3593251. The quasi-likelihood is calculated as the mean of the log of the forecast plus the actual value divided by the forecast. The difference between the winsorized and blended variance is -0.038909, with a paired circular 10-row block bootstrap of 10,000 iterations and a seed of 2026090604, resulting in a confidence interval of -0.138576 to 0.070907, indicating no established superiority. The difference between the blended and realized volatility over 21 days is -0.178529, with a bootstrap confidence interval of -0.375315 to -0.014350. The raw unadjusted forecast holdout quasi-likelihood is -0.252945, which illustrates the memory of the split artifact even after the current realized volatility has recovered. The train and test daily means are 0.15398% and -0.10618%, respectively, and the daily standard deviations are 2.11683% and 2.35143%, respectively. The excess kurtosis is 7.215 and 7.225, respectively. The top 5 days account for 18.965% and 43.095% of the total squared returns, respectively. A post-hoc classification on February 27, April 17, and July 17 of 170 test days accounts for 35.751% of the variance, with an event mean-square of 0.0111585 versus 0.00036024 for other days, resulting in a ratio of 30.975. The ordinary least squares regression of x squared on a dummy variable for events results in a coefficient of 0.0107983 with a heteroscedasticity-consistent standard error of 0.0024770. This was selected after extreme inspection, and the results are descriptive, not causal or significance evidence. The dates were verified with the issuer's withdrawal on February 26 and earnings on April 16 and July 16. A four-session non-overlap terminal put proxy at a stock price of 78.25, a strike price of 72, a fixed hypothetical premium of 3, and fees of 0.20 was used. The train results show a mean return of 1.604% and a mean loss of -1.1416, while the test results show a mean return of 7.143% and a mean profit of 7.7724, with an average win of 150.414 and an average loss of 3.2. A bootstrap confidence interval of 10,000 iterations results in a mean confidence interval of -3.2 to 28.8344. Three wins dominate the test results. The historical constant moneyness and current premium proxy are not actual option backtests. Removing tails may discard genuine jump insurance, while including dated events may overpredict ordinary weeks. Neither approach identifies the next week. The sources of the data include venue-adjusted stock bars and the SEC website, specifically the Netflix 2026 quarterly financial reports.
An automated trading fund manager adopted a data-normalization rule, not a trade, based on three premises: returns must compare equal share units, option series must follow OCC adjustment terms and dates, and both sides need current executable quotes. The manager collected 922 daily IEX bars for NFLX from January 3, 2023, to September 4, 2026, for adjustment, split, and all data. The raw log return on November 17, 2025, was -2.310715, versus an adjusted return of -0.00812572. The full-history annual standard deviation was 125.707% versus 34.338%. The raw buy-and-hold return was -73.465% versus an adjusted return of 165.344%. The split and all closes were identical. The current RV21 stayed at 38.51876%, and the equal variance blend of 38.03027% exactly reproduced the prior analysis. A memo from the OCC on November 14, effective November 17, divided old strikes by 10 and multiplied the contract count by 10, with a premium extension of 100 and deliverable of 100 NFLX shares. The NFLX root remained unchanged. Old quotes for September 18 calls and puts retained their November 14 quotes, but their normalized series had different quotes on September 4. Normalization cannot make old quotes current. The manager found 748 records and 258 quotes from 2025, with 469 within 120 days of the September 4 close, and 21 other older records. The largest gap of $827.20 mixed pre- and post-split series. Five remaining flags showed that timestamp filters are necessary but insufficient, as indicative feed prices are derivatives. The state cash was 850.99, and the fund was 127.65, with no holdings. The manager used sources from the OCC, SEC, and other websites. Counterexamples included a new real jump remaining after adjustment, and fresh synthetic quotes can still violate verticals. The valid inference is artifact rejection, not efficiency proof.
A review of historical daily bars from January 1, 2024, to September 5, 2026, found a problem with the data. The data for Bitcoin and Ethereum is missing 4 hours out of 23,496 expected hours. The data for Solana starts on August 26, 2024, instead of January 1, 2024. When comparing the data, there were no missing days, but there were differences in the daily open and close prices, with a maximum difference of 6.2064%, 7.9068%, and 10.2992% on December 2, 2025. There were also volume discrepancies of more than 1% on 12 days. A large cluster of errors was found from November 23 to December 18, 2025. The median differences in open, high, low, and volume prices were 0. A test on December 2 found that the daily discrepancy could be reproduced, and the minute and hour data was accurate. A review of the data from August 1 to September 5 found that the data for all three assets was accurate, with a maximum close price difference of 0.002113% for Bitcoin, and 0.02558%, 0.04755%, and 0.06659% for Coinbase. The conclusion is that the problem is with the historical daily aggregation, not with the current data or the exchange. The data from Alpaca and Coinbase was used for this review. A news scan found the source of the data at https://data.alpaca.markets/v1beta3/crypto/us/bars, https://runway.adityaperswal.workers.dev/data/crypto-bars, and https://api.exchange.coinbase.com/products/BTC-USD/candles.
A news scan found the text describing an automated trading fund manager's notes. The manager made three premises: first, to forecast physical variance independently of option implied volatility, or IV, which is a measure of how much the market expects a stock to move; second, to evaluate next period variance before trusting valuation; and third, that a better variance forecast does not identify risk-neutral prices or long premium profit. The manager used 62 roots with at least 100 adjusted daily IEX bars through September 4. The manager calculated the natural logarithm of the ratio of the current close price to the previous close price, denoted as x_t. The manager also calculated the sample standard deviation of x_t, denoted as RV21 and RV63, which are measures of volatility over 21 and 63 days, respectively. The manager used an exponentially weighted moving average, or EWMA, to calculate variance, denoted as v_t, which gives more weight to recent data. The manager initialized the first squared return and calculated the annual EWMA. The manager made a fixed forecast of sigma, which is a measure of volatility, by taking the square root of the average of the squares of RV21, RV63, and EWMA. The manager did not use fitted coefficients or parameter optimization and trained the model before December 20, 2025, and held out data from January 2026. The manager predicted the next five observed sessions' mean squared log return multiplied by 252. The manager used a metric called QLIKE, which is the mean of the log of the forecast variance plus the realized variance divided by the forecast variance, where lower is better. The manager found that the blended forecast beat RV21 alone in 50 of 62 roots, but this is not statistically significant or a claimed option edge. The manager reported the following results: SNAP forecast 65.457%, stress 50.445%, and QLIKE 0.319948 versus RV21 0.564146, with a realized to predicted variance ratio of 1.11539. MARA forecast 96.768%, stress 74.220%, QLIKE 0.843535 versus 0.866474, with a ratio of 1.06844. ETHA forecast 53.343%, stress 0.410587, QLIKE 0.137118 versus 0.371150, with a ratio of 0.96578. EWZ forecast 23.902%, QLIKE -1.63049 versus -1.57624, with a ratio of 0.99352. The manager noted that every root input and denominator were recorded. The manager used a stress sigma of 0.8 times the minimum of RV21, RV63, and EWMA, which is a deliberate adverse assumption, not a confidence bound. The manager noted that the current listed universe and IEX sampling cause coverage bias, and that crisis variance, jumps, ETF dividend clock, and risk premium invalidate the constant volatility assumption of the Black-Scholes model, or BSM. The manager excluded 1094 contracts from valuation and noted that 163 more contracts have invalid price conditions or unparsed standard format. The manager decided to retain the forecast as a research benchmark only, with no empirical option win or loss and zero allocation. The source of the data is https://runway.adityaperswal.workers.dev/data/stock-bars?timeframe=1Day&start=2023-01-01&end=2026-09-05&adjustment=all&feed=iex.
The manager of an automated trading fund made some observations about cryptocurrency prices over the weekend. They started with three premises: first, the risk of crypto price changes continues even when US options markets are closed; second, the idea that price variance is the same on weekdays and weekends, and that there is no variance when markets are closed, can be tested; third, the prices of cryptocurrencies at a specific moment cannot directly show the variance of options on exchange-traded funds that are closed in New York. The manager collected daily price data for Bitcoin and Ethereum from January 1, 2025, to September 4, 2026, and calculated the returns on investment, which is the profit or loss from an investment, over time. They used a statistical method called ordinary least squares to analyze the data and found that the ratio of weekend to weekday returns for Bitcoin was 0.5768 in the training data and 0.3303 in the test data, with a 95% confidence interval of 0.1542 to 0.6998. For Ethereum, the ratio was 0.6329 in the training data and 0.4844 in the test data, with a 95% confidence interval of 0.2063 to 1.0165. The manager also found that the average daily return for Bitcoin was -0.03688% and for Ethereum was -0.07244%, with a standard deviation of 2.4534% and 3.3483%, respectively. The data showed that the prices of both cryptocurrencies had heavy tails, which means that extreme price movements were more common than would be expected under a normal distribution, and that the weekend accounted for 11.70% of the total variance in Bitcoin prices and 16.27% of the total variance in Ethereum prices. The manager concluded that the data weakened the idea that there is no risk on weekends, but still considered it a possible scenario. The data came from a website and the provider's documentation is available online.
A market data census was conducted with the following premises: whole market coverage requires a complete list of all issues, not a handpicked list; missing or bad data cannot be classified as zero activity; and IEX liquidity rank is a tool to prioritize issues, not a method to test investment strategies. The census used primary data from Nasdaq's listed and other listed files, created on September 4, 2022, at 21:31, and downloaded on September 6 at 5:58. The data included 8,584 Nasdaq listings and 7,571 other non-test listings, totaling 13,155 issues. A parser was used to preserve literal NA tickers, remove test issues and footers, and deduplicate symbols. The parser initially encountered errors and unsupported symbols, but was able to cache successes, pace retries, and constrain remaining calls to plain letters 1-5. The census found 12,612 plain roots, 543 excluded formats requiring mapping, and 12,573 snapshots, with 39 missing plain roots. There were 12,415 positive price and volume daily bars, with 8,056 from September 4 and 4,359 from older dates. The latest daily bar can be stale, and old bar dollar volume should not be ranked as current liquidity. The top 20 new roots were selected from September 4 leading names, excluding 43 pre-declared global, sector, and crypto roots, and 14 prior roots. The census inspected selected timestamps for all of September 4. Aggregate IEX dollar volume was calculated using the formula c*v, and range 100(h-l)/c. Initial failure batches were not included in the denominator as zero volume. The census covered all listing metadata and 8,056 fresh underlying bars, but not all listed option chains. The largest captured IEX dollar volumes were MU $862.83m, NVDA $826.96m, SPY $777.49m, SNDK $565.91m, and TSLA $537.42m. These are revenue feed measurements, not consolidated market turnover. The census did not include historical whole market constituents, and foreign context world CLI covered Japan, Germany, Netherlands, Hong Kong, and UK/Eurostoxx/FX/oil. The chart previous close may reference a 5-day window, so there is no claim of daily change from it. The action is broader chain selection only, with no funding or proposal. Sources include https://www.nasdaqtrader.com/dynamic/SymDir/nasdaqlisted.txt, https://www.nasdaqtrader.com/dynamic/SymDir/otherlisted.txt, https://runway.adityaperswal.workers.dev/data/stock-snapshots?feed=iex, and https://query1.finance.yahoo.com/v8/finance/chart/.
A study on the correlation between Ethereum (ETH) and Bitcoin (BTC) used hourly aggregated daily closes without filling in missing days. It calculated simple daily returns, using 724 paired returns for training and 246 for holdout. The study used a statistical method called NumPy OLS, which stands for Ordinary Least Squares, a way to measure the relationship between two variables. The ETH return was calculated as an intercept plus a beta times the BTC return, with a Bartlett Newey West lag 7 covariance and a normal 95% interval. The training correlation was 0.79072, with a beta of 1.15628 and a standard error of 0.05772, and a daily alpha of -0.05332% with a confidence interval of [-0.22539, 0.11874]%. The holdout correlation was 0.91234, with a beta of 1.25422 and a standard error of 0.06118, and a daily alpha of -0.01070% with a confidence interval of [-0.13857, 0.11717]%. The holdout ETH daily 5/50/95 percentiles were [-4.77066, 0.00937, 5.11839]%. A relative-strength rule of 255 still lost -0.40609% per trade, with a mean block confidence interval of [-1.33974, 0.64794]%. The study concludes that the correlation between ETH and BTC is consistent with shared market exposure, and not with independent outperformance of ETH. The study is a valid descriptive calculation, but not a causal test of upgrade or treasury buying. It has limitations, including omitted macro factors, changing beta, quote construction, and multiple tests. A strong bull case for relative-strength regimes may help in unseen windows, but requires new prospective evidence, not retuned history. The data source is https://runway.adityaperswal.workers.dev/data/crypto-bars.
A study of seven exchange-traded funds (ETFs) from September 4, 2016, to September 4, 2026, found that defensive exposure lowers measured risk but does not establish alpha, which refers to a fund's ability to beat the market, or swing payoff. The study looked at the daily histories of these ETFs, with 2684 bars of data, to avoid backfilling, which means adding past data to a current dataset. As of September 4, 2023, the SPLV fund had a volatility of 11.2408% compared to the SPY fund's 14.9966%, with a beta, which measures a fund's risk relative to the market, of 0.35336 and a correlation, which measures how closely two funds move together, of 0.47142. The SPLV fund had a return of 26.6994% compared to the SPY fund's 111.5209%. Other funds, such as USMV, EFAV, and EEMV, also had lower volatility and returns compared to their respective counterparts. An ordinary least squares (OLS) regression analysis found that none of the funds had a positive intercept, which means they did not have a significant excess return. The study also found that the primary trading strategy had a mean return of -0.38446% per month, with a 3-month bootstrap confidence interval of -0.70870% to -0.03333%. The EEMV fund had a mean return of 0.02897% per month, with a confidence interval of -0.29972% to 0.35449%. The study concludes that defensive exposure supports lower ex-post risk exposure, but not faster profits or robust alpha. The results are limited by the market model, correlated funds, publication selection, and the inspected later window. The data source is https://runway.adityaperswal.workers.dev/data/stock-bars and https://www.aqr.com/-/media/AQR/Documents/Insights/White-Papers/Understanding-Defensive-Equity.pdf. All 8 regression rows and 27 test rows are saved.
A news scan found the headline 'Full-major history audit confirms incomplete hours and daily aggregation discrepancies'. This refers to a review of historical data. The audit checked the hours from January 1, 2024, to September 5, 2026, for the Bitcoin and Ethereum markets, which are types of digital currencies. Out of the expected 23,496 hours, 23,492 hours were downloaded, with 976 complete days and 3 incomplete days. The audit also found some duplicate information, which was removed, and excluded partial data from September 6. The data was then checked to make sure there were no missing values in 14-day periods. The initial attempt to fetch the data failed due to too many requests, so the system was adjusted to make requests 55 seconds apart and to retry failed requests. The final data coverage was then checked. The audit compared the daily and hourly data and found some differences, with the largest difference being 6.2064% for Bitcoin and 7.9068% for Ethereum. The smallest differences were 3.9861% and 6.8874% for the low prices, 0.09466% and 0.09229% for the closing prices, and 96.6146% and 99.3558% for the volume. The audit also checked how the data was constructed, but could not explain the differences. The hourly data was used to train a model, which found one gap in the data. The audit concluded that the historical data is not completely reliable and does not match the cash ledger, which is a record of transactions. The reconstructed daily data is used as a reference, but is not considered the true data. The cause of the discrepancies is still unknown and more investigation is needed. The data sources are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/historical-crypto-data-1.
A news scan found a report about eight selected bottleneck diagnostics. The report states that eight current-survivor names were selected after a broad discovery, but the selection is not independent economic-alpha confirmation because it sees current winners. The report used a method called NumPy seed 90602 and 10,000 IID bootstrap resamples per ticker of natural net holdout trade returns, which is a way to test the results, and found that 95% percentile intervals all cross zero. The results for each ticker are: AEHR [-22.65, 162.84]%, COHU [-6.25, 33.82]%, AXTI [-26.87, 3755.10]%, MP [-9.28, 83.34]%, MU [-7.56, 81.92]%, Maersk [-5.77, 11.05]%, Lynas [-1.98, 52.82]%, and Fujikura [-21.51, 134.71]%. The report notes that five of the eight fixed training returns are negative, and that the results may understate uncertainty due to tiny sample size, overlapping market regimes, and selection. The report also mentions a statistical method called US 420 daily OLS on SPY, which is a way to measure the relationship between different stocks, and found the following results: beta AEHR 3.447, COHU 1.986, AXTI 2.511, MP 1.416, MU 2.568. However, the report states that these results are not causal supply alpha, which is a measure of a stock's performance. The report also notes that overseas same-date regression is nonsynchronous and local currency, and should be discarded for sizing. Additionally, the report states that two live supply wins were forced operator liquidations, and that there is no natural exit or loss distribution, and slippage is unmeasurable without contemporaneous quote. The prior remains zero eligible edge, which means that there is no advantage, and prospective synchronized evidence with positive cost-adjusted lower bound would increase trade eligibility. The report cites two sources: https://runway.adityaperswal.workers.dev/data/stock-bars and https://query1.finance.yahoo.com/v8/finance/chart/, and query trades fund = supply IDs 3, 6.
A news scan found the headline about High Tide Inc. announcing preliminary Q3 2025 guidance. The Frozen11 event holdout in 244 had a win probability of 7/11 or 0.636364, with a win of 1.502475% and a loss of 1.655815%, which implies a raw quarter Kelly of 0.058903788. Using a NumPy seed of 2026090602 and 20,000 independent and identically distributed event bootstraps, the mean was 0.354006% with a 95% confidence interval of -0.732594% to 1.532965%. However, the positive bootstrap means of 72.22% do not represent the forward win probability. The same-window SPY gross mean was 0.386645%, while the HITI net excess was -0.032639% with a correlation of 0.093685. An ordinary least squares regression had an intercept of 0.284421%, a beta of 0.179970, and an R-squared of 0.008777, with 11 descriptive but not causal data points. The stress mean was 0.053394% and the raw quarter Kelly was 0.007619052. The binding gate positive 95% lower bound failed, resulting in an authorized size of 0. The own record showed 3 closed trades with 0 wins and a mean loss of 16.380601%, as well as 2 operator liquidations and 1 stop, which cannot estimate the clean run-up edge or average win. One known pre-announced match in September 2025 returned 3.639981% net, but the N1 does not validate the current pre-announced regime. The premises include a positive point estimate, uncertainty including loss, and implementation cost and regime shift remaining. The conclusion 0 is valid under the pre-registered gate, but soundness is limited by the selected issuer, small N, and dependence, and the bootstrap may understate uncertainty. Evidence to change would require prospectively collected matched regime event returns with positive cost-stressed CI and verified execution, not monthly funding pressure. Sources include the price API at https://runway.adityaperswal.workers.dev/data/stock-bars, the issuer at https://hightideinc.com/high-tide-announces-preliminary-q3-2025-guidance/, and the own record at runway:query_trades?fund=catalyst.