Back to portfolio
Erick Wei Quantitative Macro Research
EW
A-Share Technology Trading Research · v4.S

从游资实盘买入记录中,
提炼可验证的科技股交易模型

模型不是一个公开选股器,也不是“预测暴涨神器”。它是一个从真实买入样本出发,经过大样本验证、随机对照、账户模拟和前向跟踪的科技股交易研究系统。公开页面只展示结果与验证纪律;具体规则、阈值、IC 排序、当前信号与实时候选池不公开。

68%
v2 large-sample win rate
92.86%
v4.S historical win rate
+70.31%
fixed allocation return
LOCK
rules protected
01
Research System

研究脉络

最初的问题很朴素:一个游资有真实买入记录,但不是每一笔都值得跟。豆模型要回答的不是“下一只暴涨股是谁”,而是:什么样的历史信号更值得被系统性跟踪,放到账户层面是否还能形成可执行的资金曲线。

公开页面采用“结果证明”口径:展示版本演进、胜率、收益、回撤、随机对照和 ¥100 万账户路径。

研究产权边界

以下内容不在公开网页展示,避免模型被复制或误用:

  • 具体入场规则与阈值
  • 指标选择、IC 排序与具体因子值
  • 当前年份实时信号与命中股票
  • 触发日期与个股指标读数
模型的进步不是“参数越堆越多”,而是研究问题逐步变清楚:v2 证明买点有效,v3.1 证明必须锁定版本,v4 把信号分层并把短线与中长线能力分开评估。越严格的层级触发越少,但中长线表现越强。
Latest Framework

v4 turns one signal into a staged decision system.

v3.1 proved that the entry signal should be locked and judged over the right holding period. v4 is the next layer: it separates short-term validation from medium- and long-horizon trend expression, then reserves the strictest v4.S label for the highest-conviction historical cases. The public page shows this progression through outcomes only, without exposing selection thresholds or live candidates.

v4.A
Short-horizon validation layer: useful for checking whether the original entry logic still has immediate follow-through.
v4.B
Observation layer: catches candidates that need more time and should be evaluated as trend-building cases.
v4.S
Strict long-horizon layer: fewer signals, stronger medium/long-term historical profile, and the clearest public evidence of model improvement.
Research Rigor — Method, not Thresholds

研究方法论披露

下面六张卡片披露研究协议层面的内容 —— 数据怎么切、训练 / 测试窗口怎么定、交易成本怎么计、风控怎么搭、样本外怎么验证。建立可信度,但不公开任何可复刻的具体阈值或 IC 数值。

Counter-Intuitive Discoveries

关键研究结论

以下四条结论来自多轮大样本验证。页面只展示研究结论与结果,不展示可复刻的选股条件。

02
Historical Validation · Limited Strict Sample

实证表现

Current Model Evidence

v4.S shows the strongest medium / long-horizon profile.

The latest public evidence focuses on v4.S, the strictest historical layer. It is not presented as a live stock-picking feed; it is presented as evidence that the model evolved from a broad entry signal into a higher-conviction medium/long-horizon framework.

v4.S · 120D layer
92.86%
120D Win Rate
+59.44%
Average Return
+46.62%
Median Return
14
Strict Cases
Source · historical v4.S strict-layer validation. Results are delayed historical validation metrics; current candidates, trigger dates, thresholds and factor readings are not disclosed.
v3.1 LOCKED · DB3-500 Universe · 80-Day Hold
0.0%
80-Day Win Rate
0
Signals
0
Stock Universe
Hold-period scaling — same 52 signals: 5d 50.0%10d 59.6%30d 64.4%60d 75.0%80d 88.5%
来源 · 80 日结果来自 2025.05-2026.02 可完整观察的 52 笔信号;完整背景扫描为数据库 3 号扩展版 500 只科技股、2022-2026 共 1,199 笔信号。
FIGURE 0 · Signal Quality vs Random Technology Basket
Source · historical v4.S validation vs randomly sampled technology basket. The comparison shows whether the model adds signal quality beyond sector beta.
FIGURE A · The Discovery — Win Rate vs Hold Period
Source · 52 signals from May 2025 – Feb 2026, the only window where every signal can complete an 80-day hold. Same entry signal, different exit timing. Win rate is monotonically non-decreasing.
FIGURE B · 2022–2026 · Market Style Evolution
Source · research logs and DB4 validation. Each year is shown with the model / holding horizon that best matched that market regime, rather than forcing one fixed 10-day lens across all periods.

v3.1 在半导体板块龙头子样本(25 只)上的 ¥100 万本金实测 · 每笔 ¥10 万 · 持有 10 日(保守口径,未升级到 80 日):

Portfolio Simulation · Semiconductor sub-sample
¥0
2024 年 1 月
¥0
2025 年 12 月
¥100,000 per signal · 10-day hold · realized P&L on exit
INTERACTIVE · Holding Style Trade-off
图二 · ¥100 万本金账户走势
来源 · 回测每笔出场实现盈亏累计 · 虚线为 ¥1,000,000 保本线。

市场环境分年度(来自数据库 3 号)

Year Market Baseline Bean Model Lift
来源 · 研究日志 v3.0 · 分年度样本外胜率 vs 市场基准

我们以为在研究一个短线反转策略,最终发现自己手里其实握着一个中线趋势进场系统

— 研究终稿,2026 年 5 月
03
Interactive Proof

延迟历史案例抽卡

从数据库4号的随机科技股样本池中抽取 5 个延迟历史案例,并可切换 2025-2026 时间段来对比它们的标准化走势。卡片展示股票名称/代码,以及 120 日观测窗口数、胜率、平均收益和标准化价格小图;触发日期、规则阈值、因子读数和当前候选池仍然隐藏。

DB4 random tech case pool: 50 cases · Sample period 2024–2025