Development process

From the financial-hacker.com

1) Selecting the model. Confirming it with price data

The three prerequisites for a financial model:

Has a rational basis in market structure / trader behavior
Can be expressed in a program flow or formula
Has statistical significance in real price curves

Find an algorithm that detects the inefficiency in price curves.

Do a statistic. Plot a histogram.
Compare with random walk curves or shuffled price curves. Difference should be significant.
Do NOT rely on other people‘s research! Scam is ubiquitous (-> "Elliott Waves", Rich Swannell)

2) Developing the trade algorithm

Example: Cycle strategy

Detect the dominant cycle and phase.
Generate a forerunning sine curve.
Enter short at a sine peak.
Enter long at a sine valley.
Exit on reversal or after a half-period.

3) Developing the filter algorithm

A market inefficiency normally does not exist all the time. Therefore, we need a filter for determining if the inefficiency is present or not. In most cases the filter is more important than the algorithm. Example: Cycle strategy

Measure the amplitude of the dominant cycle.
Trade only when the amplitude is above a threshold.

4) Parameter adaption ("optimizing" / "training")

If the model has "free parameters":

Find out how the strategy reacts on parameter changes.
Find the most robust parameter range ("sweet spot").
Adapt the strategy to different assets.
Adapt it to different market situations (even while live trading).

Bad ideas:

Optimizing too many parameters.
Optimizing for peaks (= brute force or genetic optimization).

Training a strategy in Zorro Optimize Strategies in Backtrader

5) Test

Test should cover all significant market periods (5-10 years).
Any parameter adaption introduce bias to the test result.
The bias renders backtests completely useless.
The solution: Testing the system with data not used for the adaption - Walk-Forward Analysis.

Walk-Forward Analysis

Roll a window over the simulation period
Separate the window in a training and test section.
Good: The test is out-of-sample and still covers most of the data.
Bad: The system depends on two more parameters.

Main performance parameters:

Wins divided by losses (Profit Factor)
Annual profit in relation to drawdown (Calmar ratio) (Drawdown must be normalized -> square root rule!)
Annual return in relation to sigma (Sharpe ratio)
Linearity of returns (R2 coefficient)

6) Reality check

7) Implementing risk and money management

Use a stop loss for eliminating negative outliers.
Do not use profit targets. (If you really want to, use a profit-lock mechanism instead).
Use an algorithm for calculating the optimal investment per portfolio component (Kelly, OptimalF, Markowitz).
Re-invest only the square root of your profits.
Supervise your system permanently and compare live results with backtest results (-> „Cold Blood Index“).

Performance metrics for strategy comparison

Metric	Example	Description
Annual return (AR)
Profit Factor (PF)		The sum of all profits divided by the sum of all losses. A neutral strategy has a value of 1. A loosing strategy is < 1. Profitable strategies are usually between 1.1 and 1.8.

Profit Factor: Gross profit divided by the absolute value of gross loss. Profit factors of 1.5 or more suggest a strong system. A trading system or method with a low profit factor could become unprofitable with just a slight change in market dynamics | | Sharpe ratio (SR) || Reward-to-variability ratio where increased variability is equated with higher risk. The higher SR the better its returns relative to the investment risk. | | R2 coefficient (R2) || Measures the linearity of the equity curve. It is compared with a line through its start and end points. An R2 of 1 | | Ulcer index (UI) || Average depth in drawdown in percent. A good strategy should haven an UI < 10%. | |||| |Pessimistic |||