Strategy #0000000060 · Machine learning and statistical learning
Artificial neural-network direction signal
Inspect inputs, a hidden unit, sigmoid output and threshold position; separate classification from future-return assumptions.
An artificial neural network weights features, applies nonlinear transformations and produces a class score. Financial time series are prone to overfitting: leakage, feature selection, class imbalance and regime change can undermine strong training results. Evaluation needs time-ordered validation and trading costs.
One neuron is instructional only; there is no training, data cleaning or out-of-sample test, and no reliable price forecast.
Three ways in
Understand the story. Then inspect the risk.
Start with a familiar situation, check the payoff, then test how the numbers change.
The idea · in two minutes
How do technical features pass through a hidden neuron into a class score and assumed trade outcome?
Leo feeds two standardised technical features into a small network: a tanh hidden unit followed by a sigmoid up-class score.
The simulator lets him change weights, biases and the decision threshold, then applies the position to a manually entered next-period return after costs.
No model is trained here; future return is an assumption. Proper research must avoid look-ahead bias, use chronological train/test splits, compare simple baselines and inspect turnover across market regimes.
- Toy network probability-up outputAdjust two inputs, hidden/output weights, class threshold, assumed future return and costs to inspect a toy network output.
- Long-signal scenario return less costOne tanh hidden unit and sigmoid output only; weights are untrained and future return unobserved, so this is neither forecast nor backtest.
- No out-of-sample observations in this demoIllustrative expiry outcome.
Illustrative research examples only. They are not live signals or a recommendation.
The mechanics · payoff and exposure
h=\tanh(w_1x_1+w_2x_2+b_h),\quad p=\sigma(vh+b_o),\quad Position=\begin{cases}1&p\geq q\\-1&p\leq1-q\\0&otherwise\end{cases}
x₁ and x₂ are standardised inputs; w₁, w₂ and b_h are hidden parameters; v and b_o are output parameters; q is the long/short threshold.
Greeks: the directional intuition
- 前視偏差
- Measures sensitivity to implied volatility.
- 模型選擇
- Measures sensitivity to implied volatility.
- 制度轉變
- Measures sensitivity to implied volatility.
h=\tanh(w_1x_1+w_2x_2+b_h),\quad p=\sigma(vh+b_o),\quad Position=\begin{cases}1&p\geq q\\-1&p\leq1-q\\0&otherwise\end{cases}
Practical context · published observations
Artificial neural-network direction signal: read the record with its caveats.
The chart uses the published figures or explicitly labelled study inputs below. A historical index or model replay is not the same as an investor’s realised account return.
只有一個 tanh 隱藏節點和 sigmoid 輸出；手動權重未訓練，後續報酬未觀測，不能當作預測或回測。 Source ↗ · Methodology ↗
Change assumptions and inspect the scenario
Inputs illustrate the strategy mechanics; they are not live quotes, forecasts or investment advice.
Variants · same family, different risk levers
Related methods still carry different exposures and failure modes.
Compare the input, position and assumptions that can fail.
| Related strategy | Relationship | What changes | New cost or risk |
|---|---|---|---|
| Multilayer feed-forward network | Deeper nonlinear mapping | Adds hidden-layer capacity. | More parameters increase overfit risk. |
| LSTM sequence model | Sequential state | Represents temporal dependencies. | Data requirements, drift and calibration risk. |
| CNN multivariate classifier | Local patterns | Treats features as a temporal array. | Feature ordering and leakage risk. |
One tanh hidden unit and sigmoid output only; weights are untrained and future return unobserved, so this is neither forecast nor backtest.
Read beyond the summary
References
Primary sources and research used to describe the strategy and its limits.