Expand model tag support: add GLM-5.1, simplify Anthropic IDs, scan tags anywhere in message
- Flink update_bars debouncing - update_bars subscription idempotency bugfix - Price decimal correction bugfix of previous commit - Add GLM-5.1 model tag alongside renamed GLM-5 - Use short Anthropic model IDs (sonnet/haiku/opus) instead of full version strings - Allow @tags anywhere in message content, not just at start - Return hasOtherContent flag instead of trimmed rest string - Only trigger greeting stream when tag has no other content - Update workspace knowledge base references to platform/workspace and platform/shapes - Hierarchical knowledge base catalog - 151 Trading Strategies knowledge base articles - Shapes knowledge base article - MutateShapes tool instead of workspace patch
This commit is contained in:
175
gateway/knowledge/trading/strategies/crypto/ann-strategy.md
Normal file
175
gateway/knowledge/trading/strategies/crypto/ann-strategy.md
Normal file
@@ -0,0 +1,175 @@
|
||||
---
|
||||
description: "An artificial neural network (ANN) strategy that forecasts short-term BTC price movements using technical indicators (EMA, EMSD, RSI) as inputs and quantile-based classification as output."
|
||||
tags: [crypto, machine-learning, ann, bitcoin, technical-analysis]
|
||||
---
|
||||
|
||||
# Artificial Neural Network (ANN) Strategy
|
||||
|
||||
**Section**: 18.2 | **Asset Class**: Cryptocurrencies | **Type**: Machine learning / Price prediction
|
||||
|
||||
## Overview
|
||||
|
||||
This strategy uses an ANN to forecast short-term movements of BTC price based on input technical indicators. Unlike equities, cryptocurrencies have no evident "fundamentals" on which to build value-based strategies, so cryptocurrency trading strategies tend to rely on trend data mining via machine learning techniques. The ANN classifies the future normalized return into quantile buckets and generates buy/sell signals accordingly.
|
||||
|
||||
## Construction / Mechanics
|
||||
|
||||
### Price and Return Normalization
|
||||
|
||||
Let `P(t)` be the BTC price at time `t`, where `t = 1, 2, ...` is measured in some units (e.g., 15-minute intervals; `t = 1` is the most recent time).
|
||||
|
||||
**Return:**
|
||||
```
|
||||
R(t) = P(t)/P(t+1) - 1 (521)
|
||||
```
|
||||
|
||||
**Serial mean return** over T₁ periods:
|
||||
```
|
||||
R_bar(t, T₁) = (1/T₁) * sum_{t'=t+1}^{t+T₁} R(t') (523)
|
||||
```
|
||||
|
||||
**Serially demeaned return:**
|
||||
```
|
||||
R_tilde(t, T₁) = R(t) - R_bar(t, T₁) (522)
|
||||
```
|
||||
|
||||
**Variance:**
|
||||
```
|
||||
[sigma(t, T₁)]² = (1/(T₁-1)) * sum_{t'=t+1}^{t+T₁} [R_tilde(t', T₁)]² (525)
|
||||
```
|
||||
|
||||
**Normalized (serially demeaned) return:**
|
||||
```
|
||||
R_hat(t, T₁) = R_tilde(t, T₁) / sigma(t, T₁) (524)
|
||||
```
|
||||
|
||||
For notational simplicity the T₁ parameter is omitted below and `R_hat(t)` denotes the normalized return. T₁ should be chosen long enough to provide a reasonable volatility estimate.
|
||||
|
||||
### Input Layer: Technical Indicators
|
||||
|
||||
**Exponential Moving Average (EMA):**
|
||||
```
|
||||
EMA(t, lambda, tau) = ((1-lambda)/(1-lambda^tau)) * sum_{t'=t+1}^{t+tau} lambda^{t'-t-1} * R_hat(t') (526)
|
||||
```
|
||||
|
||||
**Exponential Moving Standard Deviation (EMSD):**
|
||||
```
|
||||
[EMSD(t, lambda, tau)]² = ((1-lambda)/(lambda - lambda^tau)) * sum_{t'=t+1}^{t+tau} lambda^{t'-t-1} * [R_hat(t') - EMA(t, lambda, tau)]² (527)
|
||||
```
|
||||
|
||||
**Relative Strength Index (RSI):**
|
||||
```
|
||||
RSI(t, tau) = nu_+(t, tau) / [nu_+(t, tau) + nu_-(t, tau)] (528)
|
||||
|
||||
nu_±(t, tau) = sum_{t'=t+1}^{t+tau} max(±R_hat(t'), 0) (529)
|
||||
```
|
||||
|
||||
Where: `tau` is the moving average length; `lambda` is the exponential smoothing parameter (to reduce parameters, one can set `lambda = (tau-1)/(tau+1)`).
|
||||
|
||||
Typically RSI > 0.7 is interpreted as overbought; RSI < 0.3 as oversold.
|
||||
|
||||
### Input Layer Construction
|
||||
|
||||
The input layer consists of:
|
||||
- `R_hat(t)` — the current normalized return
|
||||
- `EMA(t, lambda_a, tau_a)` for `a = 1, ..., m`
|
||||
- `EMSD(t, lambda_a, tau_a)` for `a = 1, ..., m`
|
||||
- `RSI(t, tau_{a'})` for `a' = 1, ..., m'`
|
||||
|
||||
Example parameter choices (from the literature):
|
||||
- `tau_a` corresponding to 30 min, 1 hr, 3 hrs, 6 hrs (so `m = 4`)
|
||||
- `tau_{a'}` corresponding to 3 hrs, 6 hrs, 12 hrs (so `m' = 3`)
|
||||
|
||||
### Output Layer: Quantile Classification
|
||||
|
||||
The objective is to forecast which quantile the future normalized return `R_hat(t)` will belong to.
|
||||
|
||||
Let `K` be the number of quantiles. For training dataset `D_train`, compute the `(K-1)` quantile values `q_alpha`, `alpha = 1, ..., K-1`, of `R_hat(t)`, `t in D_train`.
|
||||
|
||||
Define supervisory K-vectors `S_alpha(t)`, `alpha = 1, ..., K`:
|
||||
```
|
||||
S_1(t) = 1, if R_hat(t) <= q_1
|
||||
S_alpha(t) = 1, if q_{alpha-1} <= R_hat(t) < q_alpha, for 1 < alpha < K (530)
|
||||
S_K(t) = 1, if q_{K-1} <= R_hat(t)
|
||||
S_alpha(t) = 0, otherwise
|
||||
```
|
||||
|
||||
The output layer produces a nonnegative K-vector `p_alpha(t)` of class probabilities:
|
||||
```
|
||||
sum_{alpha=1}^{K} p_alpha(t) = 1 (531)
|
||||
```
|
||||
|
||||
### Network Architecture
|
||||
|
||||
The ANN has `L` layers labeled `l = 1, ..., L`:
|
||||
- `l = 1`: input layer
|
||||
- `l = L`: output layer
|
||||
- Intermediate layers: hidden layers
|
||||
|
||||
At each layer `l`, there are `N^(l)` nodes with vectors `X_vec^(l)` having components `X_{i(l)}^(l)`:
|
||||
|
||||
**Forward propagation:**
|
||||
```
|
||||
X_{i(l)}^(l) = h_{i(l)}^(l)(Y_vec^(l)), l = 2, ..., L (532)
|
||||
|
||||
Y_{i(l)}^(l) = sum_{j(l-1)=1}^{N^(l-1)} A_{i(l)j(l-1)}^(l) * X_{j(l-1)}^(l-1) + B_{i(l)}^(l) (533)
|
||||
```
|
||||
|
||||
Where: `A_{i(l)j(l-1)}^(l)` are the weights; `B_{i(l)}^(l)` are the biases (both determined via training).
|
||||
|
||||
**Activation functions:**
|
||||
|
||||
Hidden layers use ReLU:
|
||||
```
|
||||
h_{i(l)}^(l)(Y_vec^(l)) = max(Y_{i(l)}^(l), 0), l = 2, ..., L-1 (534)
|
||||
```
|
||||
|
||||
Output layer uses softmax (ensuring probabilities sum to 1):
|
||||
```
|
||||
h_{i(L)}^(L)(Y_vec^(L)) = Y_{i(L)}^(L) * [sum_{j(L)=1}^{N(L)} Y_{j(L)}^(L)]^{-1} (535)
|
||||
```
|
||||
|
||||
ReLU fires a neuron only if `Y_{i(l)}^(l) > 0`; softmax enforces condition (531).
|
||||
|
||||
### Training: Cross-Entropy Loss
|
||||
|
||||
The error function to minimize is the cross-entropy:
|
||||
```
|
||||
E = - sum_{t in D_train} sum_{alpha=1}^{K} S_alpha(t) * ln(p_alpha(t)) (536)
|
||||
```
|
||||
|
||||
Minimized via stochastic gradient descent (SGD), which iterates until convergence.
|
||||
|
||||
### Trading Signal
|
||||
|
||||
```
|
||||
Signal = Buy, iff max(p_alpha(t)) = p_K(t) (537)
|
||||
Sell, iff max(p_alpha(t)) = p_1(t)
|
||||
```
|
||||
|
||||
The trader buys BTC if the predicted class is `p_K(t)` (the top quantile) and sells if it is `p_1(t)` (the bottom quantile). This rule can be modified — e.g., buy on top 2 quantiles and sell on bottom 2 quantiles.
|
||||
|
||||
## Return Profile / Objective
|
||||
|
||||
The strategy profits when the ANN correctly classifies the direction and magnitude of short-term BTC price movements. Returns are driven by the quality of the technical indicator signals and the ability of the trained network to generalize out-of-sample. Given BTC's high volatility, even modest directional accuracy can produce significant returns.
|
||||
|
||||
## Key Parameters / Signals
|
||||
|
||||
- `T₁`: lookback for return normalization (volatility estimation window)
|
||||
- `tau_a`: EMA/EMSD lookback periods (e.g., 30 min, 1 hr, 3 hrs, 6 hrs)
|
||||
- `tau_{a'}`: RSI lookback periods (e.g., 3 hrs, 6 hrs, 12 hrs)
|
||||
- `lambda`: exponential smoothing factor; can be set to `(tau-1)/(tau+1)`
|
||||
- `K`: number of quantile classes (e.g., K=2 for simple up/down)
|
||||
- `N^(l)`: number of nodes at each hidden layer
|
||||
- `L`: total number of layers
|
||||
- `d_1`: number of most-recent time points excluded from training data to ensure all indicators are computed on sufficient data
|
||||
|
||||
## Variations
|
||||
|
||||
- **K=2 binary classification**: simple up/down forecast; buy/sell signal directly
|
||||
- **K>2 multi-quantile**: more granular signal strength; trade only on extreme quantiles
|
||||
- **Extended indicator set**: add MACD, Bollinger Bands, volume indicators to the input layer
|
||||
- **LSTM/RNN variant**: replace feedforward ANN with recurrent architecture to better capture time-series dependencies
|
||||
|
||||
## Notes
|
||||
|
||||
The primary risk is overfitting: many free parameters (`tau_a`, `lambda_a`, `tau_{a'}`, `N^(l)`, `K`) must be chosen and the ever-present danger of overfitting various free parameters necessitates careful out-of-sample backtesting. The training dataset must exclude the most recent `d_1` time points to ensure all EMA, EMSD, and RSI values are computed using the required number of data points. This strategy is conceptually similar to the single-stock KNN trading strategy (Section 3.17) but uses ANN instead of k-nearest neighbors. No fundamental valuation of BTC is implied.
|
||||
Reference in New Issue
Block a user