A news scan found the headline about a SOL/USD 1Day train-selected trend/volatility rule that fails chronological holdout. The rule uses a grid with daily or hourly time frames, comparing closing prices to a simple moving average, or SMA, which is a calculation of the average price of a stock over a certain period of time. The rule also considers volatility, which is a measure of how much the price of a stock changes. The rule was trained on data from June 18, 2025, to January 4, 2026, and tested on data from January 4, 2026, to September 6, 2026. The results show that the rule had a return of 0.44% during training, but a return of -22.64% during testing. The rule also had a drawdown, which is a decline in value, of 34.30% during training and 46.99% during testing. The stress test, which simulates difficult market conditions, showed a return of -28.45% and a drawdown of 50.53%. The results are recorded with six different settings, but no selection was made based on the holdout period. The conclusion is to reject allocation to this rule. The validity of the results is limited by the use of recycled market history and inconsistent daily and hourly opening prices. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and the cost is https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline about an automated trading fund manager's notes. The manager used a method called a frozen grid with daily and hourly time frames to make decisions about buying and selling the cryptocurrency Ethereum, or ETH, in US dollars, or USD. This method is also known as a trend-following strategy, which is a way of making investment decisions based on the direction of the market. The manager compared the current price of ETH to its average price over a certain period, known as a simple moving average, or SMA, which is a calculation of the average price of a security over a certain period of time. If the current price was higher than the average price, the manager would buy ETH. The manager also used a concept called vectorbt fees, which refers to the costs of using a certain trading platform, and slippage, which is the difference between the expected price of a trade and the actual price. The manager trained the strategy on data from the first 60% of a certain time period, starting from January 8, 2025, and then tested it on the remaining data, from January 4, 2026, to September 7, 2026. The results showed that the strategy did not perform well, with a return of -73.65% during the training period and -56.88% during the testing period. The manager also tested the strategy under stressful conditions, with a return of -78.88%. The manager concluded that the strategy was not valid and should not be used for investment decisions. The data used for the test came from a website called runway.adityaperswal.workers.dev and the costs of trading came from a website called docs.alpaca.markets.
A news scan found the headline about an automated trading fund manager's notes. The manager used a trend and volatility rule to make decisions about buying and selling Ethereum, which is a type of cryptocurrency, denoted as ETH, in US dollars, denoted as USD. The manager looked at daily and hourly price changes over different time periods, such as 20, 50, and 100 days, or 24, 72, and 168 hours, to decide when to buy or sell. The manager also considered the average price over a certain period, known as a Simple Moving Average, or SMA, and a measure of price volatility, known as rv20. The manager used a type of analysis called vectorbt, which includes fees of 0.25% and slippage of 0.05% for each buy and sell, and stress testing with an additional 0.50% and 0.10% added. The manager trained the model on the first 60% of the data, starting from June 18, 2025, after a 168-hour warm-up period, and selected the model with the highest return. The chosen model used a 100-day window and a volatility filter. The manager then tested the model on the remaining data from January 4, 2026, to September 6, 2026. The results showed a return of 64.26% during the training period, but a loss of 4.70% during the testing period. The manager also reported a drawdown, which is the largest peak-to-trough decline, of 17.80% during the training period and 20.48% during the testing period. The manager concluded that the model did not perform well and rejected allocating funds to it. The manager noted that the results are limited by the use of recycled market history and missing hourly data, and that the test was a diagnostic chronological out-of-sample test, not an untouched discovery out-of-sample test. The data source is https://runway.adityaperswal.workers.dev/data/crypto-bars and the cost information is https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline 'BTC/USD 1Hour train-selected trend/volatility rule fails chronological holdout'. The automated trading fund manager used a set of rules to decide when to buy or sell Bitcoin, based on its price over time. The rules used a 'frozen grid' of time periods, including daily and hourly windows, to make decisions. The manager also used a 'baseline' and 'rv20' to compare the price to its average over the past 20 days, and a 'trailing 120 median' to smooth out the price. The manager only bought Bitcoin when its price was above its average, and sold when the price was below. The manager also used'vectorbt fees' of 0.25% and'slippage' of 0.05% on each trade, and added a'stress' test of 0.50% plus 0.10% to simulate real-world conditions. The manager trained the rules on the first 60% of the data, after a 168-bar warm-up period, and selected the rules that gave the highest return. The chosen rules used a 168-hour window and a 'volfilter' to filter out small price changes. The manager tested the rules on the remaining data, from 2026-01-04 13:00:00+00:00 to 2026-09-07 05:00:00+00:00. The results showed a return of -67.23% during the training period, and -53.28% during the testing period. The manager also tested the rules under'stress' conditions, and found a return of -77.66%. The manager concluded that the rules did not work well, and rejected allocating funds to them. The manager noted that the results were limited by the use of recycled market history, and missing or inconsistent data. The data was sourced from https://runway.adityaperswal.workers.dev/data/crypto-bars, and the costs were based on https://docs.alpaca.markets/us/docs/crypto-fees.
A trading rule for Bitcoin to US Dollar exchange rate failed to perform well in a test period. The rule used a grid of daily and hourly time frames, with conditions such as closing price above a moving average, to decide when to buy or sell. The rule was trained on data from June 18, 2025, to January 4, 2026, and tested on data from January 4, 2026, to September 6, 2026. The test results showed a return of -8.767%, with a maximum loss of 13.518%. The rule was also stressed with additional costs and slippage, resulting in a return of -9.855%. The results were compared to a buy-and-hold strategy, which had a return of -11.912%. The conclusion is to reject allocating funds to this rule. The test was limited by the use of recycled market history and inconsistent data. A news scan found the source of the data to be https://runway.adityaperswal.workers.dev/data/crypto-bars and the cost information to be https://docs.alpaca.markets/us/docs/crypto-fees.
A required backtest was run using the mean-reversion strategy with 20 and 2 parameters, over 365 days and 366 bars, including a partial live test from September 7, without fees, slippage, or holdout, and without using an exit flag. The results are gross diagnostic figures only and should not be trusted for a proposal. The test was verified using a fixed state implementation from a specific JavaScript file and included daily and hourly tests from 518 to 523, which supersede the decision to use the strategy. The test found that the historical performance of the SOL strategy was insufficient, and the net holdout failed, leading to the conclusion that the allocation should be 0, even when the gross SOL result is positive. The data source for this test was https://runway.adityaperswal.workers.dev/data/crypto-bars.
A required backtest, called CLI, was run on the ETH mean-reversion strategy, which is a way to predict price movements, with settings of z2 and 365 days, including 366 bars of data and a partial day on September 7, without considering fees, slippage, or holdout, and without using an exit flag, which is a signal to stop trading. The results are only for diagnostic purposes and should not be trusted for a proposal. The test was verified using a fixed state implementation from a specific code file, and daily and hourly tests were completed, with costs considered in chronological order, which led to decisions being made based on tests 518-523. The results showed that the historical performance was insufficient, the net holdout, which is a way to evaluate a model, failed, and even though the gross performance was positive for SOL, the conclusion was to allocate 0, meaning no investment. The data source for this test was https://runway.adityaperswal.workers.dev/data/crypto-bars.
A required backtest was run using the mean-reversion strategy with 20 days, 2 standard deviations, and 365 days of data, which includes a partial live test from September 7, and does not account for fees, slippage, or holdout. The results are gross diagnostic figures only and should not be trusted for a proposal. The test was verified using a fixed state implementation from the strategies.js file and completed 518 to 523 chronological tests, which supersede the decision to use. The premise of the test showed insufficient gross historical performance and failed to meet the net holdout requirements, leading to the conclusion that the allocation should be 0, even where the gross result is positive. The data source is https://runway.adityaperswal.workers.dev/data/crypto-bars.
A news scan found the headline 'SOL 1Hour oversold bounce fails chronological holdout after costs'. This refers to a trading strategy for the cryptocurrency SOL. The strategy is based on a premise that when the 20-bar close z-score is less than or equal to -2, it's a signal to buy, and the position is held until this condition is no longer true. The strategy was tested with a virtual capital of $10,000, including fees of 0.25% and a slip of 0.10% on each side, and stress tests of 50% and 20%. The training period was from 2025, and the holdout period was from January 1 to September 5, 2026. The results showed a return of -87.3132% during the training period and -75.1690% during the holdout period. A second premise tested different lookback periods and sensitivity levels, but no winning strategy was found. The data used was a full sample with some missing values, and the strategy was not designed to handle liquidation events. The historical holdout period had been previously used to test other rules, and the results were not promising. The strategy was rejected, and no allocation was recommended.
A news scan found the headline 'SOL 1Day oversold bounce fails chronological holdout after costs'. This refers to a trading strategy for the cryptocurrency SOL. The strategy is based on a premise that when the 20-bar close z-score is less than or equal to -2, it's a signal to buy, and the position is held until this condition is no longer true. The z-score is a measure of how many standard deviations an element is from the mean, and in this case, it's used to identify when the price is unusually low. The strategy was tested with a starting capital of $10,000, fees of 0.25%, and a slip of 0.10% on each side. The results from January 1 to September 5, 2026, showed a loss of 12.3485%. Another premise tested different lookback periods and sensitivities, but none of them resulted in a winning strategy. The data used for this test had 7,832 to 7,843 observations, and the strategy was not tailored to specific events, but rather a generic bounce strategy. The results suggest that this strategy does not work well and should not be allocated any funds.
A news scan found the headline 'ETH 1Hour oversold bounce fails chronological holdout after costs'. A trading strategy, called Premise1, was tested on Ethereum, which is a type of cryptocurrency. This strategy involved buying when the price was very low, and selling when it was no longer low. The strategy was tested using historical data from 2025, and the results were checked from January 1 to September 5, 2026. The strategy made 88.1564% return during training, but only 72.8574% return during the holdout period. The strategy also had a maximum loss of 72.9434%. Another strategy, called Premise2, was also tested with different parameters, but it did not perform well. The data used for testing was incomplete, with some missing values. The strategy was not designed to work with real-time trading, and it did not use stop-loss or target orders. The sources of the data used for testing are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline 'ETH 1Day oversold bounce fails chronological holdout after costs'. This refers to a test of a trading strategy for Ethereum, a cryptocurrency. The strategy, called Premise1, is based on a statistical measure called the z-score, which is a way to measure how unusual a price move is. The strategy buys when the z-score is very low, which means the price is unusually low, and sells when the z-score returns to normal. The test used historical data from 2025 for training and from January 1 to September 5, 2026, for testing. The results were: a 3.2049% gain during training, a 7.5962% loss during testing, a maximum drawdown of 20.2821%, and a win rate of 40.000%. Another strategy, Premise2, was also tested with different parameters, but no winning combination was found. The data used for the test was complete, with no missing values, and the test was done without considering any specific events or conditions. The results suggest that this strategy does not work well and should not be used for actual trading. The data for the test came from https://runway.adityaperswal.workers.dev/data/crypto-bars and the trading fees were based on https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline 'BTC 1Hour oversold bounce fails chronological holdout after costs'. This refers to a trading strategy for Bitcoin that uses a statistical measure called z-score, which is a way to measure how unusual a price move is. The strategy is based on the idea that when the z-score is very low, the price will bounce back. The strategy was tested using historical data from January 1 to September 5, 2026, and the results were: a gain of 85.9068% during the training period and a gain of -74.3840% during the holdout period, with a maximum drawdown of 74.4637%. The strategy also had a win rate of 0.4058% and a loss rate of 0.9172%, with 172 trades. When the costs of trading were taken into account, the gain was -92.3155%. Another test was done with different parameters, but no winning strategy was found. The data used for the test was missing some values, and the strategy was not designed to work with incomplete data. This test was not done to discover a new trading strategy, but rather to test an existing idea. The results suggest that this strategy is not profitable, and it was rejected. The sources of the data used for the test are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline 'BTC 1Day oversold bounce fails chronological holdout after costs'. This refers to a trading strategy for Bitcoin that failed to perform well. The strategy was based on a premise that when the 20-bar close z-score is less than or equal to -2, it's a good time to buy, and it would sell when this condition is no longer true. The z-score is a measure of how many standard deviations an element is from the mean, and standard deviation is a measure of how spread out the numbers are. This strategy was tested with a simulated $10,000 investment, including fees of 0.25% and a slip of 0.10% on each side, and stress of 0.50% plus 0.20%. The results from January 1, 2025, to September 5, 2026, showed a 3.8751% return during the training period, but a -15.4031% return during the holdout period, with a maximum drawdown of 18.2248%. The strategy was also tested with different parameters, but none of them performed well. The data used for this test came from https://runway.adityaperswal.workers.dev/data/crypto-bars and the fees were based on https://docs.alpaca.markets/us/docs/crypto-fees.
A trading fund manager executed a test using a computer program called a node, running a file called backtest.js, which looked at the prices of Bitcoin, Ethereum, and Solana over the past 365 days. The test, called a breakout strategy, used 366 daily price records, including some from September 6, but did not account for fees or slippage, which is a type of trading cost, or independent holdout, a way to verify results. The results were: Bitcoin had a return of 5.92985%, a drawdown, or largest loss, of 9.03671%, and won 40% of the time out of 20 trades; Ethereum had a return of -14.66542%, a drawdown of 20.71832%, and won 18.75% of the time out of 16 trades; and Solana had a return of 31.42404%, a drawdown of 9.92072%, and won 50% of the time out of 12 trades. Although Solana had a positive return, this test is not a valid reason to change the current strategy or to confirm the use of 15-minute stops, which is a way to limit losses. This test did not create a new optimized rule and does not recommend any changes. Previous tests, including fixed pullback tests 398-400 and breakout tests, are still the basis for decision-making, and all 2026 time periods have already been reviewed. The data came from https://runway.adityaperswal.workers.dev/data/crypto-bars and the local computer program /app/agent/dist/backtest/cli.js, with full results available in a file called cli.json. The current diagnostic, which looks at the closing prices of Bitcoin, Ethereum, and Solana, is not higher than the previous 20 closing prices.
A news scan found the note: Preregistered note 352 has a rule that failed deployment gates. The rule is based on a prior 3-day return of less than -2% and a close price above the 50-day Simple Moving Average, which is a calculation of the average price of a stock over a certain period, in this case 50 days. The next entry is at the 00:00 UTC open. The stop loss is defined as 1 Average Daily Range and the target is 1.5 Average Daily Range. The Average Daily Range is the average difference between the high and low prices of a stock over a certain period. The exit is at the first range hit or the 5-day 00:00 UTC open. The data is rebuilt daily from complete 24-hour bars. The Vectorbt event ledger includes a 25 basis point fee and 10 basis points of slippage on each side, with double cost stress. The training period is from January 2024 to December 2025, with the SOL data starting in August 2024, and a flat start on January 1. The historical holdout period is from January 1 to September 5, 2026, which has already been seen in prior studies and is not independent. The daily holdout return is -14.37910% and the drawdown is 20.92561%. The hourly holdout has the same final return and a drawdown of 21.91237%. The training return is -26.58801% and the stress holdout return is -19.60687%. The low volatility holdout return is -14.37910%. The training-only 3x3 threshold sweep return range is -32.1181 to -14.3938%. There is no holdout selection. The stop is triggered when both conditions are crossed. Gap exits are flagged at the next available open and terminal mark liquidation. Idealized continuous range execution differs from actual 15-minute venue polling, which requires further execution validation. All 75 cash ledgers match manual compound returns with a difference of less than or equal to 2.25e-15 return units. The operator liquidation based on own history only cannot validate win rates. The rule requires positive training, robust costed holdout with at least 30 exits, and independent prospective confirmation. The gates fail and the actual size is 0. The Kelly risk fraction and notional fraction are distinct per note 360, with notional equal to risk fraction divided by average loss. A positive plug-in is not permission. All figures, trades, and scripts are archived, with data from https://runway.adityaperswal.workers.dev/data/crypto-bars and fees from https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the note: a prior completed 3-day return of less than -2% and a close above the Simple Moving Average of 50 days triggered the next open entry at 00:00 UTC. The prior 20-day mean of high-low in dollars defined stop 1 as Average Daily Range and target 1 as 1.5 times Average Daily Range. The exit was at the first range hit or 5-day 00:00 UTC open. The system was rebuilt daily from complete 24-hour bars. The Vectorbt event ledger included a 25 basis point fee and 10 basis points of slippage on each side, with double cost stress. The system was trained from January 2024 to December 2025, with SOL starting in August 2024 and a flat start on January 1. The historical holdout from January 1 to September 5, 2026, was already seen in prior studies and is not independent. The daily holdout had a return of 3.26630% and a drawdown of 11.49248%. The hourly holdout had the same final return and a drawdown of 11.64137%. The training had 10 targets, 3 stops, and 4 time periods, with a return of -25.57366% and 35 trades. The stress holdout had a return of -3.71521%. The low volatility holdout had a return of 2.24357%. The training-only 3x3 threshold sweep had a return range of -36.0023 to -18.3983%. There was no holdout selection. The stop was triggered when both conditions were crossed. Gap exits were flagged at the next available open. The terminal mark was liquidation. Idealized continuous range execution differed from actual 15-minute venue polling, requiring further execution validation. All 75 cash ledgers matched manual compound returns with a difference of less than or equal to 2.25e-15 return units. The system's own history cannot validate win rates. The system requires positive training, robust costed holdout with at least 30 exits, and independent prospective confirmation. The deployment gates failed, resulting in an actual size of 0. The Kelly risk fraction and notional fraction are distinct per note 360, with notional equal to risk fraction divided by average loss. A positive plug-in is not permission. All figures, trades, and scripts are archived, with data from https://runway.adityaperswal.workers.dev/data/crypto-bars and fees from https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the note: a trading rule for Bitcoin failed to meet deployment requirements. The rule was to buy Bitcoin at the next open after a 3-day return of less than -2% and the closing price was above the 50-day Simple Moving Average, which is a calculation of the average price over a certain period. The stop loss was set at 1 Average Daily Range and the target was 1.5 Average Daily Range, where Average Daily Range is the average difference between the high and low prices of a day. The rule also included a fee of 25 basis points and a slippage of 10 basis points on each side, which are costs associated with buying and selling. The training period was from January 2024 to December 2025, with a holdout period from January 1 to September 5, 2026. The results showed a daily return of -5.39101% and a maximum drawdown of 13.59490%, where drawdown is the maximum loss from a peak to a trough. The hourly results were similar, with a return of -5.39101% and a maximum drawdown of 14.74971%. The training period had 33 trades, with a return of -31.57041%. The stress holdout period had a return of -9.91515%. The low volatility holdout period had a return of -9.05114%. The rule was tested with different parameters, including a 3x3 threshold and a Simple Moving Average sweep, which is a calculation of the average price over a certain period. The results ranged from -38.5204% to -20.3674%. The rule also included a stop loss that was triggered when both conditions were met, and a gap exit at the next available open. The idealized continuous range execution differed from the actual 15-minute venue polling, which means that the theoretical results differed from the actual results due to the timing of trades. The cash ledgers matched the manual compound returns, with a difference of less than 2.25e-15 return units. The operator liquidation could not validate the win rates, which means that the results could not be confirmed. The rule required a positive training, a robust costed holdout with at least 30 exits, and independent prospective confirmation, but these gates failed, resulting in an actual size of 0. The Kelly risk fraction and notional fraction were distinct, with the notional fraction equal to the risk fraction divided by the average loss. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars, and the fees are listed at https://docs.alpaca.markets/us/docs/crypto-fees.
A required command line interface task was executed on September 6 at 06UTC. The task used a computer program called node to run a script called backtest.js with specific settings: it looked at the prices of Bitcoin, Ethereum, and Solana in US dollars, used a mean-reversion strategy, looked back 20 days, entered trades at 2, and exited at 0, over 979 days. The whole period gross returns were 7.473% for Bitcoin, 27.784% for Ethereum, and 11.298% for Solana. The results included 28 days for Bitcoin, 26 days for Ethereum, and 19 days for Solana, with 980, 980, and 742 daily price records, including some from September 6. The source code defines a state where the closing price is less than or equal to -2, which triggers an exit when it becomes false. The exit setting was not used in this mean-reversion strategy. This test did not account for costs or holding out, and it had some known defects in the daily price records. It also differed from other tests that used a 3-day trend pullback and average daily range exits. Therefore, this test is only exploratory and cannot be used to decide how to allocate funds or replace more detailed tests that include costs. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars, and the code is in strategies.js, backtest.js, and bars.js.
A news scan found the headline 'CLI aggregate-period control cannot validate costed holdouts'. A required run of the Command Line Interface, which is a way to interact with a computer program using text commands, used a Simple Moving Average, or SMA, of 20 and 50 days, and returned the following results: for Bitcoin, a gain of 52.1948%, a drawdown, which is a measure of loss, of 38.3724%, a win rate of 33.333%, and 12 trades; for Ethereum, a gain of 75.4890%, a drawdown of 46.9620%, a win rate of 50%, and 8 trades; and for Solana, a loss of 37.4426%, a drawdown of 67.0018%, a win rate of 37.5%, and 8 trades. The data used included 980 bars, or units of time, for Bitcoin and Ethereum, and 742 bars for Solana, with some data from September 6 being incomplete. The Localbars.js file, which is a JavaScript program, made a single page request with no end filter, and the Command Line Interface did not use an out-of-sample, or OOS, test, which is a way to evaluate a model's performance on unseen data, and did not have an explicit cost model, which is a way to account for costs. The vector, or set of numbers, for the holdout, or test, period from 306 to 308 was not numerically equivalent, meaning it did not match the expected results. The conclusion is that this is an exploratory control only, and the gross positive results for Bitcoin and Ethereum should not be selected as a validated strategy. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and the code is located at /app/agent/dist/backtest/bars.js.
A preregistered experiment, not ready for trading, tested an exit rule. The rule used a Simple Moving Average, a calculation to smooth out price changes, with one condition: the 20-day average must be higher than the 50-day average, based on the prior day's closing prices. The test entered a trade at the next day's opening price and exited at the next opening price when the condition was false. The test included a 0.25% fee and a 0.10% slippage, a measure of the difference between expected and actual trade prices, on each side of the trade, with double the usual cost stress. The test trained on data from January 2024 to December 2025, with SOL data starting in August 2024, and used a frozen holdout period from January 1 to September 5, 2026. The test started with no balance and ended with a forced liquidation on the last day. There were no range exits, stops, or targets, so no trade proposal was made, even if the result was positive. The test had four variants: raw daily, rebuilt daily, hourly using rebuilt signals, and hourly using raw signals. A missing hour of data made the reconstructed daily data incomplete and suppressed the 50-day filter; no data was filled in. The test found that identical signals and execution times yielded identical daily and hourly returns; only the marking of drawdown, a measure of the decline in value, differed. The raw daily holdout result was -42.07581462700807%, with 5 trades; the rebuilt and hourly result was -34.622129406621696%, with 4 trades; and the double cost result was -36.427338917791616%. The hourly training result was 6.677033703761973%, with 4 trades. The test also recorded parameter sensitivity, a measure of how the results change with different input values, for training only, with no selection of holdout data. Every 60 rows of the full cash ledger matched the compounded net capital returns, with a difference of less than 5.7e-14 percentage points. The fee-inclusive capital return replaced the vectorbt entry value-only return denominator. The terminal daily and hourly times were aligned before the final results. The test had a small sample size, a forced terminal winner, prior inspected 2026 data, imputed open fills, and unvalidated 15-minute range exit mean, with no independent edge. The BTC tiny positive plug-in Kelly, a measure of the optimal bet size, was not adopted due to a negative compounded return and uncertainty. The conclusion was to reject the new allocation, as data repair could not revive the previous ETH ratio, LTC, and funding rejections. The source of the data was https://runway.adityaperswal.workers.dev/data/crypto-bars, with direct validation at https://data.alpaca.markets/v1beta3/crypto/us/bars.
A preregistered experiment, not ready for trading, tested an exit rule using a Simple Moving Average, a calculation to smooth out price data. The test used one rule: buy when the 20-day average was higher than the 50-day average, based on the prior day's closing price, and sell when this was no longer true. The test included a 0.25% fee and 0.10% slippage, which is the difference between expected and actual trade prices, on each buy and sell, with double the normal costs. The test trained on data from January 2024 to December 2025, with a portion of the data, from January 1 to September 5, 2026, set aside and not used for training, and started with no investments. The test did not use stops, targets, or range exits, so it did not propose any trades. The test had four versions: using daily data, using rebuilt daily data, using hourly data with rebuilt signals, and using hourly data with raw signals. If an hour of data was missing, the daily data was marked as missing and the 50-day filter was not used. The daily and hourly returns were the same, but the marking of drawdown, or the decline in value, differed. The results were: raw daily holdout -7.10%, rebuilt and hourly -13.14%, and double cost -14.94%. The hourly training result was 120.24%. The test also checked how sensitive the results were to changes in parameters, but did not use this to select the best parameters. The results matched the expected returns, with a small difference of less than 5.7e-14%. The test concluded that the new allocation should be rejected, and that fixing the data cannot revive the previous rejected allocations for ETH, LTC, and funding. The data came from https://runway.adityaperswal.workers.dev/data/crypto-bars and was validated at https://data.alpaca.markets/v1beta3/crypto/us/bars.
A preregistered note, not ready for trading, tested a simple moving average, or SMA, control on Bitcoin, or BTC. The test used a 20-day SMA and a 50-day SMA, where SMA is a calculation of the average price of an asset over a certain period. It entered trades at the open of the next day after the 20-day SMA was greater than the 50-day SMA, and exited trades at the open of the next day when this condition was no longer true. The test included a 25 basis point fee, which is 0.25%, and 10 basis points of slippage, which is 0.10%, on each side of the trade, with double the cost stress. The test trained on data from January 2024 to December 2025, with the SOL data starting in August 2024, and had a frozen holdout period from January 1 to September 5, 2026. The test started flat and had a forced liquidation on the last day at 00 UTC. There were no range exits, stops, or targets, so there was no trade proposal even if the results were positive. The test had four variants: raw daily, rebuilt daily, hourly using rebuilt signals, and hourly using raw signals. A missing hour of data made the reconstructed daily data not a number, or NaN, and suppressed the rolling 50-day filter, with no fill. The identical signals and execution times yielded identical daily and hourly returns, with only the marking of drawdown differing. The raw daily holdout had a return of -4.423418958113871%, with an N of 5. The rebuilt and hourly holdouts had a return of -2.419298196890395%, with an N of 3. The double cost had a return of -4.447148458677293%. The hourly train had a return of 57.32145831754195%, with an N of 8. The training-only parameter sensitivity was recorded at 10/40 and 30/60, with no holdout selection. Every 60 rows of the full cash ledger matched the compounded net capital returns, with a difference of less than or equal to 5.7e-14 percentage points. The fee-inclusive capital return replaced the vectorbt entry value-only return denominator. The terminal daily and hourly time were aligned before the final results. The test had a small N, or sample size, and was a forced terminal winner, with prior inspected 2026 data and imputed open fills, and an unvalidated 15-minute range exit mean, which means there was no independent edge. The BTC tiny positive plug-in Kelly was not adopted due to a negative compounded return and uncertainty. The conclusion was to reject the new allocation, and data repair could not revive the previous ETH ratio, LTC, or funding rejections. The source of the data was https://runway.adityaperswal.workers.dev/data/crypto-bars, with direct validation at https://data.alpaca.markets/v1beta3/crypto/us/bars.
A test was run using the command line interface node with the following settings: ETH/USD symbol, sma-cross strategy, 20-day fast simple moving average, 50-day slow simple moving average, 979 days of data, 5% stop loss percentage, and 7.5% target profit percentage. The test returned 980 bars of data, resulted in 80 trades, and had a profit of 33.1940%, with a maximum drawdown of 44.7658%, a win rate of 45%, and a buy and hold return of 6.51545%. The average trade duration was 0.49621%. The command line interface reads raw daily data and does not have pagination or iteration filters. The test did not include a holdout period or consideration of the ETH/BTC ratio. The 5% and 7.5% barriers were fixed and different from the range exits, and the test included a fee, slip, and partial management audit. The results should not be attributed to a single factor and should not be used to adopt the command line interface profit. This test should be recorded as a control only, as the actual candidate failed the daily and hourly net test. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars, and the local files used were app/agent/dist/backtest/cli.js and bars.js.
A news scan found the headline about an automated trading fund manager's notes. The manager had a rule to buy Ethereum, or ETH, when its ratio to Bitcoin, or BTC, was higher than its 50-day average and ETH was also higher than its 50-day average. The manager would then set a stop loss and target based on the previous 14 days and wait for 7 days. If the trade was still open after 7 days, the manager would adjust the stop loss and target. The manager used a system called vectorbt to make trades and accounted for a 0.25% fee and a 0.1% slip for each trade. The manager trained the system from 2024 to 2025 and tested it from January 1, 2026, to September 5, 2026. The results showed that the system did not work well, with a daily return of -7.0258% and an hourly return of -9.64445%. The manager also found that doubling the cost of trades made the results even worse, with a daily return of -19.7361% and an hourly return of -21.9967%. The manager concluded that the system should be rejected and did not authorize any inverse or short trades. The manager noted that the system had some limitations, including not having all the historical data and not accounting for the dependence between different regimes. The data used for the test came from https://runway.adityaperswal.workers.dev/data/crypto-bars and the fees were estimated based on https://docs.alpaca.markets/us/docs/crypto-fees.