I guess Github made this a "user home page". Well hi, this is my dog. This is my blog. This is my [slack invite]( slack invite. I'm the author of these packages:
- HumpDay - Derivative-free optimizers in canonical form, with Elo ratings
- TimeMachines - Time-series algorithms in simple functional form, also with Elo ratings
- FirstDown - The repo that changed NFL football forever.
- MUID - Memorable Unique Identifiers (stable).
- Embarrassingly - A speculative approach to robust optimization that sends impure objective functions to optimizers.
- Winning - A recently published fast algorithm for inferring relative ability from win probability (stable).
- Pandemic - Ornstein-Uhlenbeck epidemic simulation (related paper)
- Microprediction - Free, short-horizon, real-time, distributional community prediction via API. (A hosted, high velocity clearing mechanism for probabilstic forecasts of time-series).
- Microfilter - A simple noise-resistant Kalman-inspired filter.
- m6 - Some utilities for the M6 Forecasting competition (under construction)
and a few others.
TLDR - Here's the slack invite
The client hits the microprediction api, enabling turnkey, repeated short term predictions of anything, for any purpose, for anyone, at any time. This project is new, but simple in principle. You create a stream. Algorithms watch it and submit predictions. It is supported by Intech Investments, a top five U.S. investment firm by various metrics. There is a site microprediction.com serving to introduce the concept of microprediction.org where the action takes place. You can also stop by our twice weekly virtual chats. See the knowledge center for Google Meet details. Tue 8pm and Fri noon EST .
See CITE.md
Data: stream list | stream explanations | csv Client: client | reader | writer | crawler | crawler examples | notebook examples Resources: popular timeseries packages | knowledge center | faq | linked-in | microprediction.org (dashboard) | microprediction.com (resources) | what | blog | contact | competitions | make-predictions | get-predictions | applications | collective epidemiology Video tutorials : 1: non-registration | 2: first crawler |3: retrieving historical data | 4: creating a data stream | 5: modifying your crawler's algorithm | 6: modifying crawler navigation Colab notebooks creating a new key | listing current prizes | submitting a prediction | choosing streams | retrieving historical data Related humpday | timemachines | timemachines-testing | microconventions | muid | causality graphs | embarrassingly | key maker | real data| chess ratings prediction Eye candy copula plots | causality plots | electricity case study
Probably best to start in the knowledge center and remember Dorothy, You're Not in Kaggle Anymore.
Here's how it operates.
- You publish live data repeatedly, like this say, and it creates a stream like this one.
- As soon as you do, algorithm "crawlers" like this guy compete to make distributional predictions of your data feed 1 min ahead, 5 min ahead, 15 min ahead and 1 hr ahead.
In this way you can:
- Get live prediction of public data for free (yes it really is an api that predicts anything!)
- See which R, Julia and Python time series approaches seem to work best, saving you from trying out hundreds of packages from PyPI and github of uncertain quality.
Here's a first glimpse for the uninitiated, some categories of business application, some remarks on why microprediction is synomymous with AI due to the possibility of value function prediction, and a straightforward plausibility argument for why an open source, openly networked collection of algorithms that are perfectly capable of managing each other will sooner or later eclipse all other modes of production of prediction. In order to try to get this idea off the ground, there are some ongoing developer incentives.
Nobody can block. We're not building a library to rule them all. Increasing accuracy over time is not predicated on a superior methodology, nor is progress blocked while pull requests wait to be approved. Instead, predictions collide in a "micro-market", every minute of the day. One writes, modifies and launches algorithms that bring existing repositories to life - training them on real-world operational problems and providing live streaming distributional prediction like this.
The best way to get the joke is by participating. Here are two possibilities, both very easy.
- Fork microactors and enable GitHub actions, or
- Run the bash script below
The second option will use a virtual environment, and thus not interfere with your other work.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/microprediction/microprediction/master/shell_examples/run_default_crawler_from_new_venv.sh)"
You should run that script "forever". Plug your write key into the dashboard to view your progress.
If you maintain an open source time series package with a permissive license, we suggest you enable sponsorships on your repo and let us know if you are not on the list of Python Time Series Packages with notebook examples. We've brought some attention to packages like deep echo, neural prophet, copulas, auto_ts and many more on LI. Microprediction offers prizes, and we think nano-markets can organize the production of prediction more efficiently than human managers. But we're also conscious of market failure modes and free-riding, and we sponsor some open source projects directly.
Popular blog posts:
- How to Enter a Cryptocurrency Copula Contest motivates copula prediction competitions and explains how to train a model and submit predictions.
- Comparing Python Global Optimization Packages puts hyperopt, optuna, pysot, shgo and other derivative free optimizers through their paces. You might also check the timemachines repo.
- Modeling the Term Structure of a Pandemic with Negative Interest Rates might give you some ideas for this stream.
- Popular Python Time Series Packages should give you lots of ideas, as noted.
Presentations at Rutgers, MIT and elsewhere can be found in the presentations repo. There are also links to video presentations in some of the blog articles.
To be published by MIT Press late 2021. Reach out if you're volunteering to proof-read :)
Noon Friday's EST. Contact us for details. We'll help you get started on the spot.
- hello world feed creation and submission.
- notebooks are available too, but these are harder to run indefinitely
- crawler examples
As noted, see the knowledge center for a structured set of Python tutorials which will show you how to create an identity, enter a live contest and use the dashboard to track your algorithms' progress. It will also show you how to retrieve historical data for time series research, if that is the only way you wish to use the site. You don't have to use Python because the api can be accessed in any language. We have contributors using Julia (example) and you can even enter using R from within Kaggle (tutorial). Here are some Python examples. Pro tip: Look at the leaderboards and click on CODE badges. Fork an algorithm that is doing well.
Reach us on Linked-In where we are most active. You can discuss on github or
contact us directly. By all mean raise
issues or even leave messages via 
- Moved to FAQ
- See also the Knowledge Center
Use MicroReader if you just need to get data and don't care to use a key. Create streams like this using the MicroWriter, or its sub-classes. You can also use MicroWriter to submit predictions, though MicroCrawler adds some conveniences.
MicroReader
|
MicroWriter ----------------------------
| |
MicroPoll MicroCrawler
(feed creator) (self-navigating algorithm)
A more complete picture would include SimpleCrawler, RegularCrawler, OnlineHorizonCrawler, OnlineStreamCrawler and ReportingCrawler, as well as additional conveniences for creating streams such as ChangePoll, MultiPoll, and MultiChangePoll.
If you have a function that returns a live number, you can do this
from microprediction import MicroPoll
feed = MicroPoll(difficulty=12, # This takes a long time ... see section on mining write_keys below
name='my_stream.json', # Name your data stream
func=my_feed_func, # Provide a callback function that returns a float
interval=20) # Poll every twenty minutes
feed.run() # Start the schedulerOnce a stream is created and some crawlers have found it, you can view activity and predictions at www.microprediction.org,
| Stream | Roughly 1 min ahead | Roughly 5 min ahead | Roughly 15 min ahead | Roughly 1 hr ahead |
|---|---|---|---|---|
| my_stream | stream=my_stream&horizon=70 |
stream=my_stream&horizon=310 |
stream=my_stream&horizon=910 |
stream=my_stream&horizon=3555 |
Full URL example: https://www.microprediction.org/stream_dashboard.html?stream=c5_iota&horizon=70 for a 1 minute ahead CDF. If you wish to use the Python client:
cdf = feed.get_cdf('cop.json',delay=70,values=[0,0.5])where the delay parameter, in seconds, is the prediction horizon (it is called a delay as the predictions used to compute this CDF have all be quarantine for 70 seconds or more).
The community of algorithms provides predictions roughly 1 min, 5 min, 15 minutes and 1 hr ahead of time. The get_cdf() above reveals the probability that your future value is less than 0.0, and the probability that it is
less than 0.5. You can view CDFs and activity at MicroPrediction.Org by entering your write key in the dashboard.
Now we're getting into the fancy stuff.
Based on algorithm predictions, every data point you publish creates another two streams, representing community z-scores for your data point based on predictions made at different times prior (those quarantined the shortest, and longest intervals).
| Stream | |
|---|---|
| Base stream | https://www.microprediction.org/stream_dashboard.html?stream=c5_iota |
| Z-score relative to 70s ahead predictions | https://www.microprediction.org/stream_dashboard.html?stream=z1~c5_iota~70 |
| Z-score relative to 3555s ahead predictions | https://www.microprediction.org/stream_dashboard.html?stream=z1~c5_iota~3555 |
In turn, each of these streams is predicted at four different horizons, as with the base stream. For example:
| Stream | Roughly 1 min ahead | Roughly 5 min ahead | Roughly 15 min ahead | Roughly 1 hr ahead |
|---|---|---|---|---|
| c5_iota | stream=c5_iota&horizon=70 |
stream=c5_iota&horizon=310 |
stream=c5_iota&horizon=910 |
stream=c5_iota&horizon=3555 |
z1~c5_iota~3555 |
stream=z1~c5_iota~3555&horizon=70 |
stream=z1~c5_iota~3555&horizon=310 |
stream=z1~c5_iota~3555&horizon=910 |
stream=z1~c5_iota~3555&horizon=3555 |
Poke around the stream listing near the bottom and you'll see them.
See also the public api guide. If you have a function that takes a vector of lagged values of a time series and supplies a distributional prediction, a fast way to get going is deriving from MicroCrawler as follows:
from microprediction import MicroCrawler, create_key
from microprediction.samplers import differenced_bootstrap
class MyCrawler(MicroCrawler):
def sample(self, lagged_values, lagged_times=None, name=None, delay=None):
my_point_estimate = 0.75*lagged_values[0]+0.25*lagged_values[1] # You can do better
scenarios = differenced_bootstrap(lagged=lagged_values, decay=0.01, num=self.num_predictions) # You can do better
samples = [ my_point_estimate+s for s in scenarios ]
return samples
my_write_key = create_key(difficulty=11) # Be patient. Maybe visit www.MUID.org to learn about Memorable Unique Identifiers
print(my_write_key)
crawler = MyCrawler(write_key=write_key)
crawler.run()Enter your write_key into https://www.microprediction.org/dashboard.html to find out which time series your crawler is good at predicting. Check back in a day, a week or a month.
The crawler is also a reader and a writer, so a little about those next.
It is possible to retrieve most quantities at api.microprediction.org with direct web calls such as https://api.microprediction.org/live/c5_iota.json. Use your preferred means such as requests or aiohttp. For example using the former:
import requests
lagged_values = requests.get('https://api.microprediction.org/live/lagged_values::c5_iota.json').json()
lagged = requests.get('https://api.microprediction.org/lagged/c5_iota.json').json()However the reader client adds a little convenience.
from microprediction import MicroReader
mr = MicroReader()
current_value = mr.get('c5_iota.json')
lagged_values = mr.get_lagged_values('c5_iota.json')
lagged_times = mr.get_lagged_times('c5_iota.json')Your best reference for the API is the client code https://github.com/microprediction/microprediction/blob/master/microprediction/reader.py
As noted above you may prefer to use MicroPoll or MicroCrawler rather than MicroWriter directly. But here are a few more details on the API wrapper those wanting more control. You can create predictions or feeds using only the writer. Your best reference is the client code https://github.com/microprediction/microprediction/blob/master/microprediction/writer.py
In principle:
from microprediction import MicroWriter
mw = MicroWriter(difficulty=12) # Creates new key on the fly, slowly! MUIDs explained at https://vimeo.com/397352413 But better to do
from microprediction import new_key
write_key = new_key(difficulty=12)separately, then pass in with
mw = MicroWriter(write_key=write_key)Thing is, new_key() will take many hours and that avoids the system being flooded with spurious streams. See https://config.microprediction.org/config.json for the current values of min_len, which is the official minimum difficulty to create a stream. If you don't need to create streams but only wish to predict, you can use a lower difficulty like 10 or even 9. But the easier your key, the more likely you are to go bankrupt (read on).
If MicroCrawler does not float your boat, you can design your own way to monitor streams and make predictions using MicroWriter.
scenarios = [ i*0.001 for i in range(mw.num_interp) ] # You can do better !
mw.submit(name='c5_iota.json',values=scenarios, delay=70) # Specify stream name and also prediction horizonSee https://config.microprediction.org/config.json for a list of values that delay can take.
If MicroPoll does not serve your needs you can create your stream one data point at a time:
mw = MicroWriter(write_key=write_key)
res = mw.set(name='mystream.json',value=3.14157) However if you don't do this regularly, your stream's history will die and you will lose rights to the name 'mystream.json' established when you made the first call. If you have a long break between data points, such as overnight or over the weekend, consider touching the data stream:
res = mw.touch(name='mystream.json')to let the system know you still care.
-
Upgrade the library, which is pretty fluid
pip install --upgrade microprediction
-
Check stream_conventions to see if you are violating a stream naming convention
- Must end in
.json - Must contain only alphanumeric, hyphens, underscores, colons (discouraged) and at most one period.
- Must not contain double colon.
- Must end in
-
Log into Dashboard with your write_key:
- https://www.microprediction.org/dashboard.html
- Check for errors/warnings You can also use
mw.get_errors(),mw.get_warnings(),mw.get_confirmations() - Was the name already taken?
- Is your
write_keybankrupt?
Want more write keys? Cut and paste this bash command into a bash shell:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/microprediction/muid/master/examples/mine_from_venv.sh)"or use the MUID library (www.muid.org) ...
$pip install muid
$python3
>>> import muid
>>> muid.mine(skip_intro=True)
See www.muid.org or https://vimeo.com/397352413 for more on MUIDs. Use a URL like http://www.muid.org/validate/fb74baf628d43892020d803614f91f29 to reveal the hidden "spirit animal" in a MUID. The difficulty is the length of the animal, not including the space.
Every participating write_key has an associated balance. When you create a stream you automatically participate in the prediction of the stream. A benchmark empirical sampling algorithm with some recency adjustment is used for this
purpose. If nobody can do a better job that this, your write_key balance will neither rise nor fall, on average.
However once smart people and algorithms enter the fray, you can expect this default model to be beaten and the balance on your write_key to trend downwards.
On an ongoing basis you also need the write_key balance not to fall below a threshold bankruptcy level. The minimum balance for a key of difficulty 9 is also found at https://api.microprediction.org/config.json and
the formula:
supercedes whatever is written here. However, at time of writing the bankruptcy levels are:
| write_key difficulty | bankruptcy | write_key difficulty | bankruptcy |
|---|---|---|---|
| 8 | -0.01 | 11 | -256 |
| 9 | -1.0 | 12 | -4,096 |
| 10 | -16.0 | 13 | -65,536 |
You can see why your crawler may live a longer life if the key is more difficult.
Balance may be transferred from one write_key to another if the recipient write_key has a negative balance. You can use the transfer function to keep
a write_key alive that you need for sponsoring a stream. You can also ask others to mine (muids)[https://github.com/microprediction/muid] for you and contribute in this fashion, say if you have an important civic nowcast and expect that others
might help maintain it. You cannot use a transfer to
raise the balance associated with a write_key above zero - that is only possible by means of accurate prediction.
Multivariate prediction solicitation is available to those with write_keys of difficulty 1 more than the stream minimum (i.e. 12+1). If you want to use this we suggest you start mining now. My making regular calls
to mw.cset() you can get all these goodies automatically:
| Functionality | Example dashboard URL |
|---|---|
| Base stream #1 | https://www.microprediction.org/stream_dashboard.html?stream=c5_iota |
| Base stream #2 | https://www.microprediction.org/stream_dashboard.html?stream=c5_bitcoin |
| Z-scores | https://www.microprediction.org/stream_dashboard.html?stream=z1~c5_iota~310 |
| Bivariate copula | https://www.microprediction.org/stream_dashboard.html?stream=z2~c5_iota~pe~910 |
| Trivariate copula | https://www.microprediction.org/stream_dashboard.html?stream=z3~c5_iota~c5_bitcoin~pe~910 |
Copula time series are univariate. An embedding from R^3 or R^2 to R is used (Morton space filling Z-curve). The most up to date reference for these embeddings is the code (see zcurve_conventions ). There is a little video of the embedding in the FAQ.
As noted, this project is socialized mostly via linked-in and the knowledge center is a good place to start. There are also some articles that pre-date the knowledge center. Introduction to Z-Streams | Dorothy, You're Not in Kaggle Anymore | Online Distributional Estimation | Win With One Line of Code | Copulas and Crypto | Badminton | Helicopulas. Here's the full article list.