FreqAI is a module designed to automate a variety of tasks associated with training a predictive machine learning model to generate market forecasts given a set of input features.
Features include:
* **Self-adaptive retraining**: retrain models during [live deployments](#running-the-model-live) to self-adapt to the market in an unsupervised manner.
* **Rapid feature engineering**: create large rich [feature sets](#feature-engineering) (10k+ features) based on simple user-created strategies.
* **High performance**: adaptive retraining occurs on a separate thread (or on GPU if available) from inferencing and bot trade operations. Newest models and data are kept in memory for rapid inferencing.
* **Realistic backtesting**: emulate self-adaptive retraining with a [backtesting module](#backtesting) that automates past retraining.
* **Modifiability**: use the generalized and robust architecture for incorporating any [machine learning library/method](#building-a-custom-prediction-model) available in Python. Eight examples are currently available, including classifiers, regressors, and a convolutional neural network.
* **Smart outlier removal**: remove outliers from training and prediction data sets using a variety of [outlier detection techniques](#outlier-removal).
* **Crash resilience**: store model to disk to make reloading from a crash fast and easy, and [purge obsolete files](#purging-old-model-data) for sustained dry/live runs.
* **Automatic data normalization**: [normalize the data](#feature-normalization) in a smart and statistically safe way.
* **Cleaning of incoming data**: handle NaNs safely before training and prediction.
* **Dimensionality reduction**: reduce the size of the training data via [Principal Component Analysis](#reducing-data-dimensionality-with-principal-component-analysis).
* **Deploying bot fleets**: set one bot to train models while a fleet of [follower bots](#setting-up-a-follower) inference the models and handle trades.
The user provides FreqAI with a set of custom *base* indicators (the same way as in a typical Freqtrade strategy) as well as target values (*labels*).
FreqAI trains a model to predict the target values based on the input of custom indicators, for each pair in the whitelist. These models are consistently retrained to adapt to market conditions. FreqAI offers the ability to both backtest strategies (emulating reality with periodic retraining) and deploy dry/live runs. In dry/live conditions, FreqAI can be set to constant retraining in a background thread in an effort to keep models as up to date as possible.
**Test data** - a subset of the historic data that is used to evaluate the performance of the model after training. This data does not influence nodal weights within the model.
The normal Freqtrade install process will ask the user if they wish to install FreqAI dependencies. The user should reply "yes" to this question if they wish to use FreqAI. If the user did not reply yes, they can manually install these dependencies after the install with:
The table below will list all configuration parameters available for FreqAI, presented in the same order as `config_examples/config_freqai.example.json`.
| `freqai` | **Required.**<br> The parent dictionary containing all the parameters for controlling FreqAI. <br>**Datatype:** Dictionary.
| `startup_candles` | Number of candles needed for *backtesting only* to ensure all indicators are non NaNs at the start of the first train period. <br>**Datatype:** Positive integer.
| `purge_old_models` | Delete obsolete models (otherwise, all historic models will remain on disk). <br>**Datatype:** Boolean. Default: `False`.
| `train_period_days` | **Required.**<br> Number of days to use for the training data (width of the sliding window). <br>**Datatype:** Positive integer.
| `backtest_period_days` | **Required.**<br> Number of days to inference from the trained model before sliding the window defined above, and retraining the model. This can be fractional days, but beware that the user-provided `timerange` will be divided by this number to yield the number of trainings necessary to complete the backtest. <br>**Datatype:** Float.
| `identifier` | **Required.**<br> A unique name for the current model. This can be reused to reload pre-trained models/data. <br>**Datatype:** String.
| `live_retrain_hours` | Frequency of retraining during dry/live runs. <br> Default set to 0, which means the model will retrain as often as possible. <br>**Datatype:** Float > 0.
| `expiration_hours` | Avoid making predictions if a model is more than `expiration_hours` old. <br> Defaults set to 0, which means models never expire. <br>**Datatype:** Positive integer.
| `fit_live_predictions_candles` | Number of historical candles to use for computing target (label) statistics from prediction data, instead of from the training data set. <br>**Datatype:** Positive integer.
| `follow_mode` | If true, this instance of FreqAI will look for models associated with `identifier` and load those for inferencing. A `follower` will **not** train new models. <br>**Datatype:** Boolean. Default: `False`.
| | **Feature parameters**
| `feature_parameters` | A dictionary containing the parameters used to engineer the feature set. Details and examples are shown [here](#feature-engineering). <br>**Datatype:** Dictionary.
| `include_timeframes` | A list of timeframes that all indicators in `populate_any_indicators` will be created for. The list is added as features to the base asset feature set. <br>**Datatype:** List of timeframes (strings).
| `include_corr_pairlist` | A list of correlated coins that FreqAI will add as additional features to all `pair_whitelist` coins. All indicators set in `populate_any_indicators` during feature engineering (see details [here](#feature-engineering)) will be created for each coin in this list, and that set of features is added to the base asset feature set. <br>**Datatype:** List of assets (strings).
| `label_period_candles` | Number of candles into the future that the labels are created for. This is used in `populate_any_indicators` (see `templates/FreqaiExampleStrategy.py` for detailed usage). The user can create custom labels, making use of this parameter or not. <br>**Datatype:** Positive integer.
| `include_shifted_candles` | Add features from previous candles to subsequent candles to add historical information. FreqAI takes all features from the `include_shifted_candles` previous candles, duplicates and shifts them so that the information is available for the subsequent candle. <br>**Datatype:** Positive integer.
| `weight_factor` | Used to set weights for training data points according to their recency. See details about how it works [here](#controlling-the-model-learning-process). <br>**Datatype:** Positive float (typically <1).
| `indicator_max_period_candles` | The maximum period used in `populate_any_indicators()` for indicator creation. FreqAI uses this information in combination with the maximum timeframe to calculate how many data points that should be downloaded so that the first data point does not have a NaN. <br>**Datatype:** Positive integer.
| `indicator_periods_candles` | Calculate indicators for `indicator_periods_candles` time periods and add them to the feature set. <br>**Datatype:** List of positive integers.
| `stratify_training_data` | This value is used to indicate the grouping of the data. For example, 2 would set every 2nd data point into a separate dataset to be pulled from during training/testing. See details about how it works [here](#stratifying-the-data-for-training-and-testing-the-model) <br>**Datatype:** Positive integer.
| `principal_component_analysis` | Automatically reduce the dimensionality of the data set using Principal Component Analysis. See details about how it works [here](#reducing-data-dimensionality-with-principal-component-analysis) <br>**Datatype:** Boolean.
| `DI_threshold` | Activates the Dissimilarity Index for outlier detection when > 0. See details about how it works [here](#removing-outliers-with-the-dissimilarity-index). <br>**Datatype:** Positive float (typically <1).
| `use_SVM_to_remove_outliers` | Train a support vector machine to detect and remove outliers from the training data set, as well as from incoming data points. See details about how it works [here](#removing-outliers-using-a-support-vector-machine-svm). <br>**Datatype:** Boolean.
| `svm_params` | All parameters available in Sklearn's `SGDOneClassSVM()`. See details about some select parameters [here](#removing-outliers-using-a-support-vector-machine-svm). <br>**Datatype:** Dictionary.
| `use_DBSCAN_to_remove_outliers` | Cluster data using DBSCAN to identify and remove outliers from training and prediction data. See details about how it works [here](#removing-outliers-with-dbscan). <br>**Datatype:** Boolean.
| `outlier_protection_percentage` | If more than `outlier_protection_percentage` fraction of points are removed as outliers, FreqAI will log a warning message and ignore outlier detection while keeping the original dataset intact. <br>**Datatype:** float. Default: `0.3`
| `data_split_parameters` | Include any additional parameters available from Scikit-learn `test_train_split()`, which are shown [here](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html) (external website). <br>**Datatype:** Dictionary.
| `test_size` | Fraction of data that should be used for testing instead of training. <br>**Datatype:** Positive float <1.
| `model_training_parameters` | A flexible dictionary that includes all parameters available by the user selected model library. For example, if the user uses `LightGBMRegressor`, this dictionary can contain any parameter available by the `LightGBMRegressor` [here](https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMRegressor.html) (external website). If the user selects a different model, this dictionary can contain any parameter from that model. <br>**Datatype:** Dictionary.**Datatype:** Boolean.
| `n_estimators` | The number of boosted trees to fit in regression. <br>**Datatype:** Integer.
| `learning_rate` | Boosting learning rate during regression. <br>**Datatype:** Float.
| `n_jobs`, `thread_count`, `task_type` | Set the number of threads for parallel processing and the `task_type` (`gpu` or `cpu`). Different model libraries use different parameter names. <br>**Datatype:** Float.
| `keras` | If your model makes use of Keras (typical for Tensorflow-based prediction models), activate this flag so that the model save/loading follows Keras standards. <br>**Datatype:** Boolean. Default: `False`.
| `conv_width` | The width of a convolutional neural network input tensor. This replaces the need for shifting candles (`include_shifted_candles`) by feeding in historical data points as the second dimension of the tensor. Technically, this parameter can also be used for regressors, but it only adds computational overhead and does not change the model training/prediction. <br>**Datatype:** Integer. Default: 2.
| `df['&*']` | Any dataframe column prepended with `&` in `populate_any_indicators()` is treated as a training target (label) inside FreqAI (typically following the naming convention `&-s*`). The names of these dataframe columns are fed back to the user as the predictions. For example, if the user wishes to predict the price change in the next 40 candles (similar to `templates/FreqaiExampleStrategy.py`), they set `df['&-s_close']`. FreqAI makes the predictions and gives them back under the same key (`df['&-s_close']`) to be used in `populate_entry/exit_trend()`. <br>**Datatype:** Depends on the output of the model.
| `df['&*_std/mean']` | Standard deviation and mean values of the user-defined labels during training (or live tracking with `fit_live_predictions_candles`). Commonly used to understand the rarity of a prediction (use the z-score as shown in `templates/FreqaiExampleStrategy.py` to evaluate how often a particular prediction was observed during training or historically with `fit_live_predictions_candles`). <br>**Datatype:** Float.
| `df['do_predict']` | Indication of an outlier data point. The return value is integer between -1 and 2, which lets the user know if the prediction is trustworthy or not. `do_predict==1` means the prediction is trustworthy. If the Dissimilarity Index (DI, see details [here](#removing-outliers-with-the-dissimilarity-index)) of the input data point is above the user-defined threshold, FreqAI will subtract 1 from `do_predict`, resulting in `do_predict==0`. If `use_SVM_to_remove_outliers()` is active, the Support Vector Machine (SVM) may also detect outliers in training and prediction data. In this case, the SVM will also subtract 1 from `do_predict`. If the input data point was considered an outlier by the SVM but not by the DI, the result will be `do_predict==0`. If both the DI and the SVM considers the input data point to be an outlier, the result will be `do_predict==-1`. A particular case is when `do_predict == 2`, which means that the model has expired due to exceeding `expired_hours`. <br>**Datatype:** Integer between -1 and 2.
| `df['DI_values']` | Dissimilarity Index values are proxies to the level of confidence FreqAI has in the prediction. A lower DI means the prediction is close to the training data, i.e., higher prediction confidence. <br>**Datatype:** Float.
| `df['%*']` | Any dataframe column prepended with `%` in `populate_any_indicators()` is treated as a training feature. For example, the user can include the RSI in the training feature set (similar to in `templates/FreqaiExampleStrategy.py`) by setting `df['%-rsi']`. See more details on how this is done [here](#feature-engineering). <br>**Note**: Since the number of features prepended with `%` can multiply very quickly (10s of thousands of features is easily engineered using the multiplictative functionality described in the `feature_parameters` table shown above), these features are removed from the dataframe upon return from FreqAI. If the user wishes to keep a particular type of feature for plotting purposes, they can prepend it with `%%`. <br>**Datatype:** Depends on the output of the model.
### File structure
`user_data_dir/models/` contains all the data associated with the trainings and backtests.
This file structure is heavily controlled and inferenced by the `FreqaiDataKitchen()`
Notice how the `populate_any_indicators()` is where the user adds their own features ([more information](#feature-engineering)) and labels ([more information](#setting-classifier-targets)). See a full example at `templates/FreqaiExampleStrategy.py`.
The `&*_std/mean` return values describe the statistical fit of the user defined label *during the most recent training*. This value allows the user to know the rarity of a given prediction. For example, `templates/FreqaiExampleStrategy.py`, creates a `target_roi` which is based on filtering out predictions that are below a given z-score of 1.25.
FreqAI has multiple example prediction model libraries, such as `Catboost` regression (`freqai/prediction_models/CatboostRegressor.py`) and `LightGBM` regression.
However, the user can customize and create their own prediction models using the `IFreqaiModel` class.
The user is encouraged to inherit `train()` and `predict()` to let them customize various aspects of their training procedures.
The `include_timeframes` in the example config above are the timeframes (`tf`) of each call to `populate_any_indicators()` in the strategy. In the present case, the user is asking for the
`5m`, `15m`, and `4h` timeframes of the `rsi`, `mfi`, `roc`, and `bb_width` to be included in the feature set.
The user can ask for each of the defined features to be included also from
informative pairs using the `include_corr_pairlist`. This means that the feature
set will include all the features from `populate_any_indicators` on all the `include_timeframes` for each of the correlated pairs defined in the config (`ETH/USD`, `LINK/USD`, and `BNB/USD`).
In total, the number of features the user of the presented example strat has created is:
length of `include_timeframes`* no. features in `populate_any_indicators()` * length of `include_corr_pairlist`* no. `include_shifted_candles` * length of `indicator_periods_candles`
$= 3 * 3 * 3 * 2 * 2 = 108$.
Another structure to consider is the location of the labels at the bottom of the example function (below `if set_generalized_indicators:`).
This is where the user will add single features and labels to their feature set to avoid duplication of them from
various configuration parameters that multiply the feature set, such as `include_timeframes`.
Additionally, the example classifier models do not accommodate multiple labels, but they do allow multi-class classification within a single label column.
There are two ways to train and deploy an adaptive machine learning model. FreqAI enables live deployment as well as backtesting analyses. In both cases, a model is trained periodically, as shown in the following figure.
based on the user's configuration settings. Following training, the model will be used to make predictions on incoming candles until a new model is available. New models are typically generated as often as possible, with FreqAI managing an internal queue of the coin pairs to try to keep all models equally up to date. FreqAI will always use the most recently trained model to make predictions on incoming live data. If the user does not want FreqAI to retrain new models as often as possible, they can set `live_retrain_hours` to tell FreqAI to wait at least that number of hours before training a new model. Additionally, the user can set `expired_hours` to tell FreqAI to avoid making predictions on models that are older than that number of hours.
Backtesting mode requires the user to have the data pre-downloaded (unlike in dry/live mode where FreqAI automatically downloads the necessary data). The user should be careful to consider that the time range of the downloaded data is more than the backtesting time range. This is because FreqAI needs data prior to the desired backtesting time range in order to train a model to be ready to make predictions on the first candle of the user-set backtesting time range. More details on how to calculate the data to download can be found [here](#deciding-the-sliding-training-window-and-backtesting-duration).
The user defines the backtesting timerange with the typical `--timerange` parameter in the
configuration file. The duration of the sliding training window is set by `train_period_days`, whilst
`backtest_period_days` is the sliding backtesting window, both in number of days (`backtest_period_days` can be
a float to indicate sub-daily retraining in live/dry mode). In the presented example config,
the user is asking FreqAI to use a training period of 30 days and backtest on the subsequent 7 days.
This means that if the user sets `--timerange 20210501-20210701`,
FreqAI will train have trained 8 separate models at the end of `--timerange` (because the full range comprises 8 weeks). After the training of the model, FreqAI will backtest the subsequent 7 days. The "sliding window" then moves one week forward (emulating FreqAI retraining once per week in live mode) and the new model uses the previous 30 days (including the 7 days used for backtesting by the previous model) to train. This is repeated until the end of `--timerange`.
Although fractional `backtest_period_days` is allowed, the user should be aware that the `--timerange` is divided by this value to determine the number of models that FreqAI will need to train in order to backtest the full range. For example, if the user wants to set a `--timerange` of 10 days, and asks for a `backtest_period_days` of 0.1, FreqAI will need to train 100 models per pair to complete the full backtest. Because of this, a true backtest of FreqAI adaptive training would take a *very* long time. The best way to fully test a model is to run it dry and let it constantly train. In this case, backtesting would take the exact same amount of time as a dry run.
During dry/live mode, FreqAI trains each coin pair sequentially (on separate threads/GPU from the main Freqtrade bot). This means that there is always an age discrepancy between models. If a user is training on 50 pairs, and each pair requires 5 minutes to train, the oldest model will be over 4 hours old. This may be undesirable if the characteristic time scale (the trade duration target) for a strategy is less than 4 hours. The user can decide to only make trade entries if the model is less than
FreqAI stores new model files each time it retrains. These files become obsolete as new models are trained and FreqAI adapts to new market conditions. Users planning to leave FreqAI running for extended periods of time with high frequency retraining should enable `purge_old_models` in their config:
The user can include these metrics by assigning them to `dk.data['extra_returns_per_train']['my_new_value'] = XYZ` inside their custom prediction model class. FreqAI takes the `my_new_value` assigned in this dictionary and expands it to fit the return dataframe to the strategy.
The user needs to set the standard dictionary in the config so that FreqAI can return proper dataframe shapes. These values will likely be overridden by the prediction model, but in the case where the model has yet to set them, or needs a default initial value, this is the value that will be returned.
to indicate to the bot that it should not train models, but instead should look for models trained by a leader with the same `identifier`. In this example, the user has a leader bot with the `identifier: "example"`. The leader bot is already running or launching simultaneously as the follower.
The feature set created by the user is automatically normalized to the training data. This includes all test data and unseen prediction data (dry/live/backtest).
The test data is used to evaluate the performance of the model after training. If the test score is high, the model is able to capture the behavior of the data well. If the test score is low, either the model either does not capture the complexity of the data, the test data is significantly different from the train data, or a different model should be used.
Model training parameters are unique to the machine learning library selected by the user. FreqAI allows the user to set any parameter for any library using the `model_training_parameters` dictionary in the user configuration file. The example configuration file (found in `config_examples/config_freqai.example.json`) show some of the example parameters associated with `Catboost` and `LightGBM`, but the user can add any parameters available in those libraries.
where $W_i$ is the weight of data point $i$ in a total set of $n$ data points. Below is a figure showing the effect of different weight factors on the data points (candles) in a feature set.
`train_test_split()` has a parameters called `shuffle` that allows the user to keep the data unshuffled. This is particularly useful to avoid biasing training with temporally auto-correlated data.
The user can tell FreqAI to remove outlier data points from the training/test data sets using a Dissimilarity Index by including the following statement in the config:
```json
"freqai": {
"feature_parameters" : {
"DI_threshold": 1
}
}
```
Equity and crypto markets suffer from a high level of non-patterned noise in the form of outlier data points. The Dissimilarity Index (DI) aims to quantify the uncertainty associated with each prediction made by the model. The DI allows predictions which are outliers (not existent in the model feature space) to be thrown out due to low levels of certainty.
To do so, FreqAI measures the distance between each training data point (feature vector), $X_{a}$, and all other training data points:
where $d_{ab}$ is the distance between the normalized points $a$ and $b$. $p$ is the number of features, i.e., the length of the vector $X$. The characteristic distance, $\overline{d}$ for a set of training data points is simply the mean of the average distances:
$\overline{d}$ quantifies the spread of the training data, which is compared to the distance between a new prediction feature vectors, $X_k$ and all the training data:
The parameter `shuffle` is by default set to `False` to ensure consistent results. If it is set to `True`, running the SVM multiple times on the same data set might result in different outcomes due to `max_iter` being to low for the algorithm to reach the demanded `tol`. Increasing `max_iter` solves this issue but causes the procedure to take longer time.
The user can configure FreqAI to use DBSCAN to cluster and remove outliers from the training/test data set or incoming outliers from predictions, by activating `use_DBSCAN_to_remove_outliers` in the config:
DBSCAN is an unsupervised machine learning algorithm that clusters data without needing to know how many clusters there should be.
Given a number of data points $N$, and a distance $\varepsilon$, DBSCAN clusters the data set by setting all data points that have $N-1$ other data points within a distance of $\varepsilon$ as *core points*. A data point that is within a distance of $\varepsilon$ from a *core point* but that does not have $N-1$ other data points within a distance of $\varepsilon$ from itself is considered an *edge point*. A cluster is then the collection of *core points* and *edge points*. Data points that have no other data points at a distance $<\varepsilon$areconsideredoutliers.Thefigurebelowshowsaclusterwith$N =3$.
FreqAI uses `sklearn.cluster.DBSCAN` (details are available on scikit-learn's webpage [here](#https://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html)) with `min_samples` ($N$) taken as double the no. of user-defined features, and `eps` ($\varepsilon$) taken as the longest distance in the *k-distance graph* computed from the nearest neighbors in the pairwise distances of all data points in the feature set.
new candles automatically for future retrains. This means that if new pairs arrive later in the dry run due to a volume pairlist, it will not have the data ready. However, FreqAI does work with the `ShufflePairlist` or a `VolumePairlist` which keeps the total pairlist constant (but reorders the pairs according to volume).
