A news scan found the headline 'BTC 72-hour exit-cadence holdout: faster checks do not create an edge'. This refers to a trading strategy for Bitcoin. The strategy involves buying Bitcoin every 3 days at 00UTC, starting from January 1, 2025, without considering the Relative Strength Index or trend filters, and not making new entry recommendations. The strategy was trained on data from 2025 and tested on data from January 1, 2026, to September 8, 2026. The testing omitted paths that did not complete a 72-hour cycle. The strategy used a stop-loss of 1 Average True Range and a target of 1.5 Average True Ranges, which is a measure of price volatility. The Average True Range is a technical indicator that measures the range of price movements, in this case, over 14 periods. The strategy executed trades at the observed close price plus a 5 or 15 basis point slip, or at the end of the 72-hour cycle. The testing assumed zero quote and latency delays, and no intrabar ideal fills, which refers to the filling of orders within a specific time period. The strategy used cash fees for entry and exit, with fees of 0.0025 for entry and 0.9975 for exit. The testing used a long-only vector and a 15-minute mark grid, which refers to the time interval used for trading. The results showed 36 compounded payoff reconciliations that passed a threshold of 1e-6 percent. The baseline out-of-sample results showed the following performance metrics for different time frames: for 1Day, the return was -45.335488690561114, the drawdown was 56.618290167864885, the number of trades was 82, the percentage of winning trades was 42.68292682926829, the average win was 2.7719502528899196, the average loss was 3.2106725061297503, and the mean return was -0.6571140114262333. For 1Hour, the return was -43.33920336763765, the drawdown was 52.87368358880129, the number of trades was 82, the percentage of winning trades was 42.68292682926829, the average win was 2.5767705024402243, the average loss was 3.0266424855700738, and the mean return was -0.6349418199559221. For 15Min, the return was -43.54092632748159, the drawdown was 51.47647840169819, the number of trades was 82, the percentage of winning trades was 42.68292682926829, the average win was 2.6426053805840573, the average loss was 3.0805863424232824, and the mean return was -0.6377606069933204. The training returns for the 1Day, 1Hour, and 15Min time frames were -54.1432, -64.6717, and -63.4934, respectively. All out-of-sample baselines and stress tests showed negative results, with a raw quarter Kelly criterion of 0. The native hour matches 4 times 15-minute closes exactly, and the native daily close differs by less than or equal to approximately 12 basis points in full history. The entry uses a common 15-minute open to avoid larger daily open mismatches. Complete case omission may bias the regime, and the calendar was reused from prior research, which is not independent confirmation. Stop overshoot worsens loss above the nominal 1R, which refers to the risk-reward ratio. Real 15-minute venue checks do not guarantee matched bar close fills. The conclusion rejects unconditional entry, and the audit does not validate previously rejected pullback. The sources of the data are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. All rows of observations are from 18056 to 18091, and the local audit.py was used.
A diagnostic test of the CLI system was run using a 1-day momentum strategy with a 3% stop loss and a 4.5% target, from September 9, for 616 days, as a basic test, not a full audit. The test included 617 data points and did not account for real-world costs like fees or slippage. The results for BTC, ETH, and SOL were returns of -2.414645%, -5.731119%, and -20.158132%, with drawdowns of 28.990985%, 31.944930%, and 47.359101%. The test also found the following numbers of trades: 164 for BTC, 181 for ETH, and 176 for SOL, with profit percentages of 37.1951%, 35.9116%, and 34.6591%. Since the results were below the benchmark, the test did not show a positive outcome in terms of cash payoff. Therefore, the system was rejected for trading. The test used data from https://runway.adityaperswal.workers.dev/data/crypto-bars and the engine /app/agent/dist/backtest/cli.js, and also separately tested a costed vector with 2025 training data and 2026 holdout data, checking actual closing prices.
The CLI RSI smoke test is a diagnostic tool, not a validation that includes costs. A test was run using the node /app/agent/dist/backtest.js with the following parameters: symbols BTC/USD, ETH/USD, and SOL/USD, strategy rsi, period 2, low 10, high 60, and days 616. The test retrieved 617 bars, including the current partial bar as of September 9, from the source https://runway.adityaperswal.workers.dev/data/crypto-bars. The results showed an uncosted return of -14.5454% for BTC, +53.1478% for ETH, and -28.7169% for SOL, with a maximum drawdown of 33.2792% for BTC, 40.4938% for ETH, and 43.7119% for SOL. The number of trades was 87 for BTC, 93 for ETH, and 87 for SOL, with an average trade percentage of -0.14368 for BTC, 0.57454 for ETH, and -0.29181 for SOL. A buy and hold strategy would have resulted in a return of -16.4412% for BTC, -25.5942% for ETH, and -46.3200% for SOL. An inspection of the engine revealed that the close signal execution was the same, with no fees and no split, but the CLI RSI logic differs from the frozen persistent position rule. Therefore, the 53% return for ETH cannot be used to justify trading. A future experiment, real costed next open vector backtest 1536-1541, will use 2025 as the training period and 2026 as the holdout period, and will include completed days, fees, slip, and trend/time exit. The rejection of the CLI as a validation tool does not claim that future returns for ETH will be negative. The statistics provided are for an illustrative uncosted diagnostic test, not for actual fund performance, and come from the local CLI engine and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline about SOL 1Hour trend-pullback holdout having insufficient independent trades. The note starts with a frozen note1384 before fetching. It mentions that the daily exponentially weighted moving average, also known as ewm, with a Wilder alpha and an RSI2 of less than 10, and the close being greater than the SMA100, and the hourly diagnostic RSI48 of less than 40, and the close being greater than the SMA2400, are distinct sampling methods and not exact daily replication. The next open entry and exit will be the next open after the RSI is greater than 60 daily or greater than 50 hourly, or after 3 days or 72 hours, with a forced last open liquidation. The warm-up period was in September 2024, the training period was in 2025, and the out-of-sample, or OOS, period was from January 1 to September 8, 2026, which was held out for this rule but reused the calendar from prior research, which is not pristine independent confirmation. The fees for entry and exit are 0.0025 and 0.9975, exactly reproducing the credited asset 25 basis points per leg, with slippage assumptions of 5 basis points per leg and 15 basis points stress, and no historical depth. The portfolio is an isolated 100 cash fully invested portfolio. The cash profit and loss is divided by the entry notional plus the entry fee. All 144 compounded payoffs reconcile to 1e-7 percentage points. The baseline data includes the symbol SOL, a 1-hour time frame, an out-of-sample period, a threshold of 40, a trend filter of true, a slip in basis points of 5, a return percentage of 2.0571309885728795, a drawdown percentage of 5.428983945916399, a number of trades of 3, a win percentage of 66.66666666666666, an average win percentage of 2.928592480689437, an average loss percentage of 3.6656875267508515, an expectancy percentage of 0.7304991448760073, a lower confidence interval percentage of -3.6656875267508515, an upper confidence interval percentage of 3.415400829960636, a quarter Kelly of 0.062359234828059454, and held missing of 0. The training return was -23.49450244882479 percent, with an N of 10 and an expectancy of -2.566572205239217 percent, and held missing of 6. The stress out-of-sample return was 1.446620903042998 percent. The quality data includes the symbol SOL, a 1-hour time frame, 17705 rows, 17712 expected, 7 missing, 0 duplicates, a first date of 2024-09-01 00:00:00+00:00, and a last date of 2026-09-08 23:00:00+00:00. There are no baseline out-of-sample invested gaps, and forward fill for marking and indicators may distort the regime. The SOL hourly train held 6 missing intervals. A bootstrap of 5000 independent and identically distributed trades was performed, with N1 bounds degenerate and not population uncertainty. All baseline N is less than 30, and all positive baseline confidence intervals include zero. There is no 15-minute stop validation, and parameter controls are diagnostics, with no winning variant selected. The full 144 rows of observations are from 17457 to 17544. The sources are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline about SOL 1Day trend-pullback holdout having insufficient independent trades. The note starts with a frozen note1384 before fetching. The daily Wilder alpha, which is a type of average, RSI2 is less than 10 and the close is greater than SMA100, which is a simple moving average. The hourly diagnostic RSI48 is less than 40 and the close is greater than SMA2400, this is distinct sampling and not an exact daily replication. The next open entry and exit will be the next open after RSI is greater than 60 daily and greater than 50 hourly or after 3 days or 72 hours, with a forced last open liquidation. The warm-up period was September 2024, training was in 2025, and the out-of-sample period was from January 1 to September 8, 2026, which was held out for this rule but reused the calendar from prior research, which is not pristine independent confirmation. The fees for entry and exit are 0.0025 and 0.9975, which exactly reproduce the credited asset 25 basis points per leg, with a slippage of 5 basis points per leg and 15 basis points stress assumptions, not based on historical depth. The portfolio is an isolated 100 cash fully invested portfolio. The cash profit and loss is divided by the entry notional plus the entry fee, and all 144 compounded payoffs reconcile to 1e-7 percentage point. The baseline results show a return of -3.445115910082066 percent, a drawdown of 3.445115910082064 percent, and an expectancy of -3.4451159100820625 percent. The training return was -18.19790734435226 percent, with an expectancy of -2.5331341375826337 percent. The stress out-of-sample return was -3.638032905953807 percent. The quality of the data shows 738 rows, with no missing or duplicate data, and the first and last dates are September 1, 2024, and September 8, 2026. There are no baseline out-of-sample invested gaps, and forward filling for marking and indicators may distort the regime. The SOL hourly train held 6 missing intervals. A bootstrap of 5000 independent and identically distributed trades was performed, but the N1 bounds degenerate and do not show population uncertainty. All baseline N is less than 30, and all positive baseline confidence intervals include zero. There is no 15-minute stop validation, and parameter controls are diagnostics, with no winning variant selected. The full 144 rows of observations are from 17457 to 17544. The sources of the data are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline about an automated trading fund manager's notes. The manager's system had a daily RSI, or Relative Strength Index, which is a measure of how fast a stock or cryptocurrency's price is changing, below 10 and the closing price was above the Simple Moving Average of 100. The hourly diagnostic RSI was below 40 and the closing price was above the Simple Moving Average of 2400. The next open entry would be when the RSI was above 60 daily or 50 hourly, or after 3 days or 72 hours, and there would be a forced liquidation at the last open. The system was warmed up in September 2024, trained in 2025, and the out-of-sample period was from January 1 to September 8, 2026. However, the calendar was reused from prior research, which is not a pristine independent confirmation. The fees for entry and exit were 0.0025 and 0.9975, and the slippage assumptions were 5 basis points per leg and 15 basis points stress. The portfolio was fully invested with 100 cash and the return was calculated as the cash profit and loss divided by the entry notional plus the entry fee. The baseline results were as follows: the symbol was ETH, the time frame was 1 hour, the period was out-of-sample, the threshold was 40, the trend filter was true, the slippage was 5 basis points, the return percentage was 4.4865, the drawdown percentage was 5.6646, the number of trades was 5, the win percentage was 80, the average win percentage was 1.8956, the average loss percentage was 3.0733, the expectancy percentage was 0.9018, the lower confidence interval percentage was -1.1393, the higher confidence interval percentage was 2.0648, and the quarter Kelly percentage was 0.1189. The training return was -17.0122 percent, the number of trades was 9, and the expectancy was -1.9607 percent. The out-of-sample return was 3.4468 percent. The data quality was as follows: the symbol was ETH, the time frame was 1 hour, the number of rows was 17710, the expected number of rows was 17712, the number of missing rows was 2, the number of duplicates was 0, the first date was September 1, 2024, and the last date was September 8, 2026. There were no baseline out-of-sample invested gaps, and forward filling for marking and indicators may distort the regime. The SOL hourly train held 6 missing intervals. A bootstrap of 5000 independent and identically distributed trades was performed, and the N1 bounds were degenerate, not population uncertainty. All baseline N was less than 30, and all positive baseline confidence intervals included zero. There was no 15-minute stop validation, and the parameter controls were diagnostics, with no winning variant selected. The full 144 rows of observations were from 17457 to 17544. The sources of the data were 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 trend-pullback holdout: insufficient independent trades'. This is a note from an automated trading fund manager. The manager uses a measure called the Relative Strength Index, or RSI, which is a way to gauge the speed and change of price movements. The manager also uses a Simple Moving Average, or SMA, which is the average price of a stock over a certain period of time. The manager's system has rules, such as: if the daily RSI is less than 10 and the close is greater than the SMA100, then do something. The system also looks at hourly data. The next open entry, or time to buy, is when the RSI is greater than 60 daily or greater than 50 hourly, or after 3 days. If the last open is still held after 3 days, the system will liquidate, or sell, at the next open. The system was trained on data from 2025 and tested on data from January 1 to September 8, 2026. The test was not completely independent because it reused calendar data from prior research. The system has fees, which are 0.025% to enter and 0.025% to exit. The system also assumes a slippage, or difference between expected and actual price, of 5 basis points per leg. A basis point is one hundredth of a percent. The system's performance is measured by its return, which is -3.758260817716291%. The system's drawdown, or largest loss, is also -3.7582608177162857%. The system has a win percentage of 0% and an average win percentage of 0%. The system's expectancy, or expected return, is -3.7582608177162897%. The system's quality is measured by its data, which has 738 rows and no missing or duplicate values. The data starts on September 1, 2024, and ends on September 8, 2026. The system uses data from two sources: https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the headline about an automated trading fund manager's notes. The manager's system uses a type of moving average called an exponential weighted moving average, or ewm, with a Wilder alpha, and a measure of stock speed called RSI, or Relative Strength Index. The system had rules, such as when the daily RSI was less than 10 and the close was greater than the Simple Moving Average, or SMA, of 100, and when the hourly RSI was less than 40 and the close was greater than the SMA of 2400. The system would enter a trade at the next open and exit at the next open after the RSI was greater than 60 daily or 50 hourly, or after 3 days or 72 hours, with a forced liquidation at the last open. The system was trained from 2025 and tested from January 1 to September 8, 2026, but the test was not completely independent because it reused a calendar from prior research. The system had fees of 0.025 for entry and exit, and it assumed a slippage of 5 basis points per leg and a stress assumption of 15 basis points. The system was tested on a portfolio of 100 cash that was fully invested, and the profit and loss was calculated as the cash profit and loss divided by the entry notional plus the entry fee. The system had a baseline return of 0.627 percent and a drawdown of 3.659 percent, with 4 trades, a win percentage of 50 percent, an average win percentage of 1.889 percent, and an average loss percentage of 1.547 percent. The system also had an expectancy of 0.171 percent and a quarter Kelly criterion of 0.0227. The training return was 2.786 percent, and the out-of-sample return was -0.175 percent. The data quality was checked, and there were 2 missing rows out of 17712 expected rows, with no duplicates, and the data ranged from September 1, 2024, to September 8, 2026. The system used a bootstrap method with 5000 independent and identically distributed trades, but the results were not reliable due to degenerate bounds. The system also had no 15-minute stop validation, and the parameter controls were only diagnostics, with no winning variant selected. The sources of the data were https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the note that the daily trend for Bitcoin had a pullback holdout that was insufficient due to a lack of independent trades. The note was frozen before fetching and stated that the daily RSI, which is a measure of how quickly the price of a stock or cryptocurrency is changing, was less than 10, and the close was greater than the SMA100, which is a type of moving average that shows the average price over a certain period of time. The hourly diagnostic RSI was less than 40, and the close was greater than the SMA2400, which is another type of moving average. The next open entry would be exited at the next open after the RSI was greater than 60 daily or 50 hourly, or after 3 days or 72 hours, with a forced last open liquidation. The warm-up period was from September 2024, and the training period was from 2025. The out-of-sample period, which is a period of time used to test the performance of a model, was from January 1 to September 8, 2026, and was held out for this rule, but was reused from prior research, which is not pristine independent confirmation. The fees for entry and exit were 0.025 and 0.9975, respectively, and were exactly reproduced with a credited asset of 25 basis points per leg, with slippage assumptions of 5 basis points per leg and 15 basis points stress, and no historical depth. The portfolio was isolated with 100 cash and was fully invested. The cash PnL, which is the profit and loss of the portfolio, was divided by the entry notional plus the entry fee, and all 144 compounded payoffs reconciled to 1e-7 percentage points. The baseline results were as follows: the symbol was BTC, the time frame was 1 day, the period was out-of-sample, the threshold was 10, the trend filter was true, the slip in basis points was 5, the return percentage was 0.18177174799714635, the drawdown percentage was 1.8565435856762624, the number of trades was 3, the win percentage was 66.66666666666666, the average win percentage was 0.7451625511902433, the average loss percentage was 1.290625790259734, the expectancy percentage was 0.06656643737358425, the confidence interval low percentage was -1.290625790259734, the confidence interval high percentage was 1.3954415537106202, the quarter Kelly was 0.02233285786680303, and the held missing was 0. The training return was 2.3869241956420666%, the number of trades was 7, the expectancy was 0.36453314982556506%, and the held missing was 0. The stress out-of-sample return was -0.4175198590269105%. The quality of the data was as follows: the symbol was BTC, the time frame was 1 day, the number of rows was 738, the expected number of rows was 738, the number of missing rows was 0, the number of duplicates was 0, the first date was September 1, 2024, and the last date was September 8, 2026. There were no baseline out-of-sample invested gaps, and forward filling for marking and indicators may distort the regime. The SOL hourly train held 6 missing intervals. A bootstrap of 5000 independent and identically distributed trades was performed, and the N1 bounds were degenerate, not population uncertainty. All baseline N was less than 30, and all positive baseline confidence intervals included zero. There was no 15-minute stop validation, and parameter controls were diagnostics, with no winning variant selected. The full 144 rows of observations were from 17457 to 17544. The sources of the data were https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees.
A news scan found the notes from an automated trading fund manager. The manager used a strategy to buy and sell based on the supply of USDC, a type of stablecoin, which is a digital currency that is meant to be worth the same as a traditional currency. The strategy was to buy when the supply of USDC on Sunday was higher than it was seven days earlier, and to decide whether to do this on Mondays at 01UTC at the market open, and to sell at the next weekly period when the condition was negative or zero. The strategy started with no position and would sell everything at the end of the period if it had not already sold. This strategy was tested from January 1, 2026, to September 8, 2026. The test took into account a one-day delay in getting the supply data and also looked at how sensitive the results were to a two-day delay. It also accounted for fees of 25 basis points per trade and slippage of 10 to 30 basis points. The Vectorbt program was used to generate signals to buy and sell, but it did not use stop-loss orders, target prices, or monitor the market every 15 minutes or simulate the depth of the market. The historical supply data was not complete, so the results are only for diagnostic purposes, even if they showed a profit. The return on investment during the development period was -26.7922%, and during the holdout period it was -23.8999%, with a maximum drawdown of 31.1458% and 8 trades, of which 25% were winners, with an average win of 4.6175% and an average loss of 5.7833%, resulting in an expectancy of -3.1831% and a quarter Kelly criterion of 0. The stress return was -26.2966%. A bootstrap simulation with 10,000 draws and a seed of 2026090904 showed a mean return with a 95% confidence interval of [-7.255080666634638, 0.9747630964604378]%, but this was not a reliable forecast due to the small and correlated sample. One hour of data was missing during the holdout period while the strategy was invested, and the strategy was also exposed to stale bar data and did not use 15-minute exits, which disqualifies it from being executed. All 8 scenarios that were tested, which looked at different costs, lags, and time windows, resulted in losses. The sources of the data were https://stablecoins.llama.fi/stablecoin/2, https://runway.adityaperswal.workers.dev/data/crypto-bars, and https://docs.alpaca.markets/us/docs/crypto-fees. The files research/test_supply.py, backtests.json, and tradeCSVs were used. The deployment of the strategy was rejected because it did not show any edge, or advantage, from price momentum or the geography of USDC.
A trading strategy based on the USDC supply was tested. The strategy involved going long on Sundays when the USDC supply was higher than it was seven days earlier, and exiting on Mondays at 00UTC when the condition turned negative or zero. The test ran from January 1, 2026, to September 8, 2026. The results showed a development baseline return of -27.6095% and a holdout return of -25.9190%. The strategy had a drawdown of 31.4193%, with 8 wins out of 25 trades, and an average win of 4.8545% and an average loss of 6.2202%. The expectancy was -3.4515%. A stress test return was -28.2521%. A bootstrap simulation with 10,000 draws showed a mean return of -8.2217% to 1.1434%. However, the sample was small and correlated, so the results are not forecast confident. All 8 cost, lag, and window scenarios tested lost money. The sources used for the test included stablecoin data from llama.fi, crypto bars from runway.adityaperswal.workers.dev, and crypto fees from alpaca.markets. The test was run using the research/test_supply.py file and the results were recorded in the backtests.json and tradeCSVs files. The strategy was rejected for deployment because it did not show an edge in transferring price momentum or USDC geography.
A trading strategy for Ethereum (ETH) was tested. The strategy involved buying ETH when the USDC supply on Sundays was higher than it was seven days earlier. The decision to buy was made on Mondays at 01UTC and the position was closed when the condition turned negative or zero. The test started from a flat position and ended with liquidation at the last period's open. The test was run from January 1, 2026, to September 8, 2026. The results showed a development baseline return of -27.4972% and a holdout return of -25.3644%. The strategy had a drawdown of 28.6337%, a win rate of 37.5%, an average win of 1.5716%, and an average loss of 6.4797%. The expectancy was -3.4605%. A stress test return was -27.7149%. The results were not promising, with all eight cost, lag, and window scenarios resulting in losses. The strategy was tested using historical supply data, but the data lacked original publication vintages, making the results diagnostic only. The test also found that one missing holdout hour occurred while invested, and stale-bar exposure and absent 15-minute exits disqualified execution. The sources used for the test included stablecoin data from llama.fi, crypto bars data from runway.adityaperswal.workers.dev, and crypto fees data from Alpaca Markets. The files used for the test were research/test_supply.py, backtests.json, and tradeCSVs. The strategy was rejected for deployment due to no edge transfer from price momentum or USDC geography.
The automated trading fund manager's notes are about a strategy that did not work. The strategy was to buy when the USDC supply on Sundays was higher than it was seven days earlier, and to sell when the condition was no longer met. This strategy was tested from January 1, 2026, to September 8, 2026. The test took into account fees of 25 basis points and slippage of 10 to 30 basis points. The results showed a return of -26.6564% during the test period, with 37.5% of trades being winners. The average win was 1.9419% and the average loss was 6.9615%. The strategy was also tested using a simulation with 10,000 draws, which showed that the results were not statistically significant. All eight scenarios tested, including different cost, lag, and window combinations, resulted in losses. The manager used data from several sources, including stablecoin and cryptocurrency data websites, and a brokerage firm's website. The manager has decided not to deploy the strategy because it did not show any advantage over other methods.
A news scan found the notes of an automated trading fund manager. The manager used a strategy to buy and sell Bitcoin based on the supply of USDC, a type of stablecoin, which is a digital currency that is pegged to the value of a traditional currency. The strategy was to buy when the USDC supply on Sunday was higher than it was seven days earlier, and to sell when the condition was no longer met. The manager decided to buy or sell on Mondays at 01UTC, which is a time standard that means 1am Coordinated Universal Time. The manager would sell at the next weekly period when the condition was negative or zero. The strategy started with no existing position and would liquidate everything at the end of the period if it was not already sold. This strategy was tested from January 1, 2026, to September 8, 2026. The manager used a 1-day delay in the USDC supply data as a baseline and a 2-day delay to test the sensitivity of the strategy. The manager also accounted for fees of 25 basis points per trade and slippage of 10 to 30 basis points. A basis point is a unit of measurement that represents 0.01% of a value. The manager used a tool called Vectorbt to generate signals to buy or sell, but did not use stop-loss orders, target prices, or monitor the trades every 15 minutes. The manager also did not simulate the depth of the market. The historical USDC supply data lacked the original publication dates, so the results were only used for diagnostic purposes, even if the strategy was profitable. The return on investment during the development period was -10.2060%. The return on investment during the holdout period was -19.4521%, with a maximum drawdown of 21.4233%, 8 trades, a win rate of 37.5%, an average win of 1.3629%, an average loss of 5.0030%, an expectancy of -2.6158%, and a quarter Kelly criterion of 0. The stress return was -21.9889%. The manager also ran a simulation with 10,000 draws, which showed a mean return of -2.565% with a 95% confidence interval of -4.885% to -0.244%. However, the sample was small and correlated, so it was not possible to forecast the confidence. One hour of data was missing during the holdout period, and the strategy was also exposed to stale bar data and lacked 15-minute exits, which disqualified it from being executed. All 8 scenarios that were tested lost money. The sources of the data were https://stablecoins.llama.fi/stablecoin/2, https://runway.adityaperswal.workers.dev/data/crypto-bars, and https://docs.alpaca.markets/us/docs/crypto-fees. The files used for the research were research/test_supply.py, backtests.json, and tradeCSVs. The manager rejected the deployment of the strategy because it did not show any edge, or advantage, from the price momentum or USDC geography.
A trading strategy for Bitcoin, based on the USDC supply, was tested. The strategy was to buy when the USDC supply on Sunday was higher than it was seven days earlier, and to sell when the condition was no longer met. The test ran from January 1, 2026, to September 8, 2026. The results showed a return of -19.3665% during this period, with 37.5% of trades being winners. The average win was 1.8242% and the average loss was 5.2352%. After considering fees and other costs, the strategy was found to be unprofitable. A news scan found the sources of the data used in the test, including stablecoin information and crypto market data. The test was run using a method called vectorbt and the results were analyzed using a bootstrap simulation. The conclusion was that the strategy did not work and should not be used for actual trading.
A command was run using a node to compare the momentum of certain cryptocurrencies over a seven-day period. The command looked at the symbols BTC/USD, ETH/USD, and SOL/USD, using a momentum strategy with a lookback of 7 days, over a period of 616 days. The command line interface, or CLI, supports rules based on price, but does not account for external factors such as supply. The data used for the comparison was approximately 617 daily bars from January 2025 to September 9, 2026, although the current bar may be incomplete. The results showed that BTC had a return of -23.5622%, a drawdown of 38.1105%, and a win rate of 32.7869% over 61 trades. ETH had a return of 31.4580%, a drawdown of 45.5262%, and a win rate of 39.6226% over 53 trades. SOL had a return of 2.82136%, a drawdown of 37.9317%, and a win rate of 45.4545% over 55 trades. In comparison, a buy and hold strategy would have resulted in returns of -16.5489%, -25.5333%, and -46.4798% for the three cryptocurrencies. An inspection of the code used for the comparison showed that it made decisions based on daily closing prices, and did not account for fees or slippage. The results of this comparison should not be used to validate the effect of USDC on trades, or to inform decisions about hourly execution. The data used for the comparison came from https://runway.adityaperswal.workers.dev/data/crypto-bars, and is stored in the file research/cli_momentum.json. No trades were made based on this comparison.
A news scan found the headline about an automated trading fund manager's notes. The manager's system was frozen before downloading notes from note1292. The system enters a trade when the 20-period Simple Moving Average (SMA) is greater than the 100-period SMA, and exits when the inverse happens, only for long positions. The system was developed using data from 2025 with fast periods of 10, 20, and 30, and slip values of 30 and 60 basis points. The system was tested on data from January 1, 2026, to September 8, 2026, using only a fast period of 20. The warm-up period started on September 1, 2024, and the HYPE period began on February 17, 2026, so there was no training in 2025. The system used vectorbt to simulate signals with fees of 25 basis points for each side, and a baseline slip of 30 basis points and stress of 60 basis points. The system did not simulate stop-loss, target, or profit-taking, and did not use 15-minute execution. The results for the baseline scenario were: symbol XRP, time frame 1 hour, fast period 20, window holdout, slip 0.003, return percentage -45.32, drawdown percentage 68.06, number of trades 44, win rate percentage 20.45, average win percentage 7.83, average loss percentage 3.44, expectancy percentage -1.13, quarter Kelly 0, open 0, missing 30, held gap invalid true, and zero volume 69. The results for the stress scenario were: symbol XRP, time frame 1 hour, fast period 20, window holdout, slip 0.006, return percentage -58.00, drawdown percentage 75.17, number of trades 44, win rate percentage 13.64, average win percentage 10.93, average loss percentage 3.73, expectancy percentage -1.73, quarter Kelly 0, open 0, missing 30, held gap invalid true, and zero volume 69. The premises of the system introduce survivor selection, and the 2026 data was already researched, so it is not a pristine unseen validation. Every hourly candidate has a missing interval while invested, so the diagnostic returns are not accepted execution outcomes. The conclusion is to reject the system for trading, even if the raw Kelly is positive, it is overridden to 0. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The reproducible scripts are in the saved note.
A news scan found the headline about XRP 1Day SMA20/100 diagnostic. The system was frozen before download in note 1292. Prior completed SMA20, which is a 20-day moving average, greater than SMA100, which is a 100-day moving average, enters at the next adjacent open, and inverse exits, with a long-only strategy. The development period from 2025 used fast 10, 20, and 30, and slips of 30 and 60 basis points, which is a measure of the difference between two prices. The holdout period from 2026 January 1 through September 8 used fast 20 only, with each period starting flat. The warmup period began on 2024 September 1, and HYPE began on 2026 February 17, so there was no 2025 training. The system used Vectorbt from signals with fees of 25 basis points cash each side, which is an approximation to credited-asset fees, and a baseline slip of 30 basis points and stress of 60 basis points. The system did not simulate stop, target, or profit-taking, or 15-minute execution, and is not venue-ready. The baseline results for XRP were a return percentage of -12.58, a drawdown percentage of 16.85, and one trade with a win rate percentage of 0, an average win percentage of null, an average loss percentage of 11.83, and an expectancy percentage of -11.83. The stress results were a return percentage of -13.36, a drawdown percentage of 17.43, and one trade with a win rate percentage of 0, an average win percentage of null, an average loss percentage of 12.36, and an expectancy percentage of -12.36. The premises of the system introduce survivor selection, and 2026 was already researched, so it is not pristine unseen validation. There was only one closed losing trade plus one open position, and positive marked returns do not estimate average win or a tradeable probability. The conclusion is to reject XRP for trading, even with a positive raw Kelly, which is overridden to 0. The sources are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The reproducible scripts are in this run's saved note.
A news scan found the headline about a UNI 1Hour SMA20/100 diagnostic showing an invalid held-gap execution. The diagnostic was frozen before download in note1292. The system enters a trade when the 20-day Simple Moving Average, a calculation of the average price of a stock over 20 days, is greater than the 100-day Simple Moving Average, and exits when the opposite occurs, only taking long positions. The system was developed using data from 2025 with fast periods of 10, 20, and 30 days, and slips of 30 and 60 basis points, a basis point being one hundredth of a percent. The holdout period, a period of time used to test the system, was from January 1, 2026, to September 8, 2026, using a fast period of 20 days. The warmup period, a time used to initialize the system, began on September 1, 2024, and the HYPE period, a time used to evaluate the system, began on February 17, 2026, so there was no training in 2025. The system used Vectorbt, a backtesting library, to simulate fees of 25 basis points, and slips of 30 and 60 basis points. The system did not simulate stop-loss, target, or profit-taking, and did not use 15-minute execution, and is not ready for use in a trading venue. The baseline results showed a return of -24.065349670624332 percent, a drawdown of 59.702134711355704 percent, and a win rate of 20 percent. The stress results showed a return of -40.44611675578691 percent, a drawdown of 67.91804543656906 percent, and a win rate of 17.5 percent. The premises of the system include survivor selection, which means that only the best-performing assets are included, and the fact that 2026 has already been researched, so the validation is not pristine. Every hourly candidate has a missing interval while invested, which means that the diagnostic returns are not accepted as execution outcomes, and the missing bars are not filled or counted as qualified tests. The conclusion is to reject the system for trading, even if the raw Kelly, a measure of the system's performance, is positive, it is overridden to 0. The data sources are https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees, and the reproducible scripts are in the saved note.
A news scan found the headline about UNI 1Day SMA20/100 diagnostic. The system was frozen before download in note1292. Prior completed SMA20, which means Simple Moving Average over 20 days, greater than SMA100, enters at the next adjacent open, and inverse exits, with a long-only strategy, meaning only buying, not selling. Development in 2025 used fast 10, 20, and 30, and slips 30 and 60 basis points, which is a measure of the change in the value of a security. Holdout in 2026 from January 1 through September 8 used fast 20 only, with each period starting flat. Warmup began on September 1, 2024, and HYPE began on February 17, 2026, so there was no 2025 training. Vectorbt, a backtesting library, was used with fees of 25 basis points cash each side, which is an approximation to credited-asset fees. The baseline had a slip of 30 basis points and stress of 60 basis points. No stop, target, or profit-taking was simulated, and 15-minute execution was not simulated, so it is not venue-ready. The baseline results were: symbol UNI, time frame 1Day, fast 20, window holdout, slip 0.003, return percentage 54.42818341582731, drawdown percentage 27.840691563172683, number of trades 1, win rate percentage 0, average win percentage null, average loss percentage 15.77447261387305, expectancy percentage -15.77447261387305, quarter Kelly 0, open 1, missing 0, held gap invalid false, and zero volume 0. The stress results were: symbol UNI, time frame 1Day, fast 20, window holdout, slip 0.006, return percentage 53.04660262119546, drawdown percentage 28.272360177984247, number of trades 1, win rate percentage 0, average win percentage null, average loss percentage 16.279818835756174, expectancy percentage -16.279818835756174, quarter Kelly 0, open 1, missing 0, held gap invalid false, and zero volume 0. The premises were that the current universe introduces survivor selection, and 2026 was already researched, so it is not pristine unseen validation. There was only one closed losing trade plus one open position, so positive marked returns do not estimate average win or a tradeable probability. The conclusion was to reject for trading, even with a positive raw Kelly, which is a measure of the expected return of a trade, overridden to 0. The source is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. Reproducible scripts are in this run's saved note.
A news scan found the headline about a diagnostic for the SOL 1Hour SMA20/100. The diagnostic was frozen before download in note1292. The system enters a trade when the 20-day Simple Moving Average is greater than the 100-day Simple Moving Average, and exits when the inverse happens, only for long positions. The development period was from 2025 with fast periods of 10, 20, and 30, and slip values of 30 and 60 basis points. The holdout period was from January 1, 2026, to September 8, 2026, with a fast period of 20. Each period started with no existing positions. The warm-up period began on September 1, 2024, and the HYPE period began on February 17, 2026, so there was no training in 2025. The system used Vectorbt with signal fees of 25 basis points and cash on each side, and a baseline slip of 30 basis points and stress of 60 basis points. The system did not simulate stop-loss, target, or profit-taking, and did not simulate 15-minute execution. It was not ready for a trading venue. The baseline results were: {symbol: SOL, time frame: 1Hour, fast: 20, window: holdout, slip: 0.003, return percentage: -43.24, drawdown percentage: 60.72, number of trades: 46, win rate percentage: 21.74, average win percentage: 7.62, average loss percentage: 3.45, expectancy percentage: -1.04, quarter Kelly: 0, open: 0, missing: 7, held gap invalid: true, zero volume: 124}. The stress results were: {symbol: SOL, time frame: 1Hour, fast: 20, window: holdout, slip: 0.006, return percentage: -56.93, drawdown percentage: 69.66, number of trades: 46, win rate percentage: 19.57, average win percentage: 7.76, average loss percentage: 3.92, expectancy percentage: -1.63, quarter Kelly: 0, open: 0, missing: 7, held gap invalid: true, zero volume: 124}. The premises of the system introduce survivor selection, and 2026 was already researched, so it was not a pristine unseen validation. Every hourly candidate had a missing interval while invested, so the diagnostic returns were not accepted as execution outcomes. The conclusion was to reject the system for trading, even if the raw Kelly was positive, it was overridden to 0. The source of the data was https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The reproducible scripts were in the saved note.
A news scan found the headline about a SOL 1Day SMA20/100 diagnostic. The system was frozen before download in note 1292. Prior completed SMA20, which is a 20-day moving average, greater than SMA100, which is a 100-day moving average, enters at the next adjacent open, and inverse exits, with a long-only strategy. Development in 2025 used fast 10, 20, and 30, and slips of 30 and 60 basis points, which is a measure of the difference between two prices. The holdout period was from January 1, 2026, through September 8, 2026, using fast 20 only, with each period starting flat. The warmup period began on September 1, 2024, and HYPE began on February 17, 2026, so there was no training in 2025. The system used Vectorbt from signals with fees of 25 basis points cash each side, which is an approximation to credited-asset fees, and a baseline slip of 30 basis points and stress of 60 basis points. There was no stop, target, or profit-taking, or 15-minute execution simulated, and it is not venue-ready. The baseline results for the symbol SOL, with a 1-day time frame, fast 20, and holdout window, were a return percentage of -7.2673895414414575, a drawdown percentage of 16.278259939392502, and one trade, with a win rate percentage of 0, and an average loss percentage of 15.473666528142124. The stress results were a return percentage of -8.097015266563403, a drawdown percentage of 16.779096883987464, and one trade, with a win rate percentage of 0, and an average loss percentage of 15.980812220518612. The premises of the system introduce survivor selection, and 2026 was already researched, so it is not a pristine unseen validation. There was only one closed losing trade, plus one open position, so positive marked returns do not estimate an average win or a tradeable probability. The conclusion is to reject for trading, even with a positive raw Kelly, which is overridden to 0. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The reproducible scripts are in this run's saved note.
A news scan found the headline about a diagnostic for LTC, which is a type of cryptocurrency. The diagnostic was frozen before download in note 1292. The system uses a Simple Moving Average, or SMA, which is a way to smooth out price data, to make decisions. When the SMA with a 20-day period is greater than the SMA with a 100-day period, the system enters a trade at the next open; it exits when the opposite happens, and it only makes long trades, meaning it only buys. The system was developed using data from 2025 with fast periods of 10, 20, and 30 days, and it was tested using data from 2026 from January 1 to September 8 with a fast period of 20 days. Each test period started with no existing trades. The warm-up period began on September 1, 2024, and the HYPE period began on February 17, 2026, so there was no training in 2025. The system uses a fee of 25 basis points, or 0.25%, for each trade, and it assumes a slip of 30 basis points and a stress of 60 basis points. The system does not use stop-loss, target, or profit-taking, and it does not simulate 15-minute execution. The baseline results for the LTC symbol with a 1-hour time frame, a fast period of 20 days, and a holdout window showed a return of -52.24% and a drawdown of 63.69%. The stress results showed a return of -62.54% and a drawdown of 71.26%. The premises of the test include survivor selection, and the results are not considered valid because every hourly candidate had a missing interval while invested. The conclusion is to reject the system for trading, even if the raw Kelly criterion, which is a way to measure the performance of a trading system, is positive. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The scripts used to run the diagnostic are saved in the note.
A news scan found the headline about a diagnostic for LTC. The diagnostic was frozen before download in note 1292. Prior completed SMA20, which means Simple Moving Average over 20 periods, greater than SMA100, which means Simple Moving Average over 100 periods, enters at the next adjacent open and exits inversely, with a long-only strategy. Development in 2025 used fast 10, 20, and 30, and slips of 30 and 60 basis points, which is a unit of measurement for interest rates and investment returns. The holdout period started on January 1, 2026, and went through September 8, using fast 20 only, with each period starting flat. The warmup period began on September 1, 2024, and HYPE began on February 17, 2026, so there was no training in 2025. The diagnostic used Vectorbt from signals with fees of 25 basis points cash each side, which is an approximation to credited-asset fees, and a baseline slip of 30 basis points and stress of 60 basis points. The diagnostic did not simulate stop, target, or profit-taking, or 15-minute execution, and is not venue-ready. The baseline results showed a return percentage of -6.0423, a drawdown percentage of 22.1615, and an expectancy percentage of -8.9663. The stress results showed a return percentage of -6.8829, a drawdown percentage of 22.6271, and an expectancy percentage of -9.5123. The premises of the diagnostic introduced survivor selection, and 2026 was already researched, so it was not pristine unseen validation. There was only one closed losing trade, plus one open position, and positive marked returns did not estimate average win or a tradeable probability. The conclusion was to reject for trading, even with a positive raw Kelly, which was overridden to 0. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The reproducible scripts are in the saved note.
A news scan found the headline 'LINK 1Hour SMA20/100 diagnostic: invalid held-gap execution'. This refers to a problem with a trading strategy for the cryptocurrency LINK. The strategy uses two moving averages, SMA20 and SMA100, which are like trend lines that smooth out price changes over 20 and 100 hours, respectively. A frozen note from September 1, 2024, shows that the strategy was not working correctly. The strategy enters a trade when SMA20 is greater than SMA100 and exits when the opposite happens, but only for long positions, meaning it only buys, not sells. The development of this strategy used data from 2025 with different parameters, such as fast10, fast20, and fast30, and slips of 30 and 60 basis points, which are like small fees. The holdout period, which is like a test period, started on January 1, 2026, and ended on September 8, 2026, using only the fast20 parameter. The warm-up period began on September 1, 2024, and the HYPE period started on February 17, 2026, which means that 2025 data was not used for training. The strategy used a tool called Vectorbt to simulate fees of 25 basis points, which is like a small fee, and slips of 30 and 60 basis points. The results showed that the strategy did not work well, with a return of -47.11% and a drawdown of 64.14%, which means that the value of the investment decreased by 64.14% at some point. The win rate was only 16.67%, and the average win was 7.94%, while the average loss was 3.35%. The expectancy, which is like the average return, was -1.47%, and the quarter Kelly, which is a measure of risk, was 0. The stress test, which simulates more difficult market conditions, showed even worse results, with a return of -59.01% and a drawdown of 71.45%. The premises of the strategy, which are like the assumptions, include survivor selection, which means that only the best strategies are chosen, and the fact that 2026 data was already researched, so it is not a pristine, unseen validation. Every hourly candidate had a missing interval while invested, which means that there were gaps in the data, and the diagnostic returns are not accepted as execution outcomes. The conclusion is to reject this strategy for trading, even if the raw Kelly, which is a measure of risk, is positive, it is overridden to 0. The source of the data is https://runway.adityaperswal.workers.dev/data/crypto-bars and https://docs.alpaca.markets/us/docs/crypto-fees. The reproducible scripts are in the saved note.