# Strategy to label stock prices

Suppose I have the prices of an action XXXX at the second for three years. I want to use the data to train my Deep Neural Network model for standard day trading purposes (i.e. High Frequency Trading). However, I have to label each second of the data file with "-1", "0", "1" for sale order, do nothing or a buy order. I don't want to label each row manually because it will take a year to do this task. How could I parse the data with python to label each row by "-1" if it is a good time to sell, "0" for doing nothing or "1" for a buy order? For a time X, we know precisely the past prices and the future prices, so it could be doable to label all of them I figure.

Be aware that for a standard trading, we have to take into account the transaction fees. In my case, I pay \$4.95 for a buy and \$4.95 for a sell. For instance, if I buy 100 shares at $1, then in order not to lose money, I must at least cover the loss related transaction costs. Hence, the selling price much be at least \$1 + (\$9.90/100) = \$1.099 ~ \$1.1 Another important aspect is that my market making strategy is order flow. So I don't want to trade on low price changes. In fact, I want the machine assists me to trade at the beginning. So I need the machine to trade as a trader normal would do on a normal day trading.  ... 2018-04-02 17:59:33.643 0.8400 2018-04-02 18:08:58.808 0.8420 2018-04-02 18:09:50.003 0.8400 2018-04-02 18:09:50.003 0.8400 2018-04-02 18:12:27.183 0.8400 2018-04-02 18:16:04.064 0.8400 2018-04-02 18:25:07.933 0.8450 2018-04-02 18:25:41.331 0.8450 2018-04-02 18:26:54.375 0.8400 2018-04-02 18:26:54.376 0.8300 2018-04-02 18:37:39.056 0.8250 2018-04-02 18:38:23.336 0.8250 2018-04-02 18:38:23.339 0.8250 2018-04-02 18:51:15.689 0.8449 2018-04-02 19:04:41.140 0.8449 2018-04-02 19:17:38.169 0.8450 2018-04-02 19:26:59.634 0.8450 2018-04-02 19:27:47.407 0.8450 2018-04-02 19:37:19.775 0.8449 2018-04-02 19:37:29.165 0.8449 2018-04-02 19:41:05.906 0.8449 2018-04-02 19:54:18.236 0.8400 2018-04-02 19:54:20.944 0.8449 2018-04-02 19:54:20.946 0.8449 2018-04-02 19:56:32.848 0.8412 2018-04-02 19:56:32.851 0.8412 2018-04-02 19:57:04.354 0.8400 2018-04-02 19:57:04.355 0.8400 2018-04-02 19:58:02.221 0.8400 2018-04-02 19:59:31.622 0.8400  ## UPDATE The mid-price at time$t$is denoted by $$p_t = \frac{s_t^{a,1} + s_t^{b,1}}{2}.$$ This mid-price can evolve in minimum increments of half a tick but is almost always observed to move at increments of a tick over time intervals of a millisecond or less. In our feature set, each limit order book update is recorded as an observation. Each observation is labelled bases on whether the mid-price will increase, decrease or remain over a horizon$h$: $$Y_t = \Delta p^t_{t+h},$$ where$\Delta p^t_{t+h}$is the forecast the discrete mid-price changes from time$t$to$t+h$, given measurement of the predictors up to time$t$. The forecasting horizon$h$can be chosen to represent a fixed number of events or can be a fixed time interval. This definition is from A High Frequency Trade Execution Model for Supervised Learning (https://arxiv.org/pdf/1710.03870.pdf). • hidden-markov-models are a well-worn path to inducing labels on sequential data, but the labeling might not coincide with the labels you have in mind (but... that's not quite clear, either). Since the ultimate goal is to trade profitably, this might be framed as a reinforcement learning task. stats.stackexchange.com/questions/377595/… – Sycorax Jul 14, 2021 at 18:11 • I guess the transaction costs are near zero now Aug 3, 2023 at 0:38 ## 2 Answers You’re trying to model an opinion, not a fact. If you want to model the sign of the return, that is a fact: the asset gained, lost, or held its value. However, you want to model what kind of trade to execute. That is an opinion, and if you buy or sell, someone else must have the opposite opinion to be selling to sell you the asset or buy it from you. Consequently, you need to come up with some kind of strategy to express your opinion about what to do. Once you do that, you can apply it to the rows of your data, but you don’t have to do that to make predictions about the times to buy, sell, and hold, as you already have that assignment rule. Just apply that rule, perhaps in a loop or with some vectorized code. Unless you want to phrase this in terms of being a game where the goal is to profit given some rules of the game and then train using reinforcement learning (which does not require the kind of labeling you describe), I do not see this as a machine learning problem. You just need to develop a rule about when to buy, sell, and hold, and then you apply that rule. You haven't stated what the criteria are for assigning labels of -1, 0, or 1. Also, if you are trading on the second time scale, then the price change per bar will be quite low. If the price change is +0.01 for a bar, and the asset price is \$1 per share (\$100 per trade for 100 shares), then then your profit will be \$1 from that trade, which won't cover the \\$4.95 commission. If you are trading 100 shares per second, where on Earth are you going to find a brokerage that will process your Sell trade alerts/signals for a given second and settle the trade in milliseconds, so that if the next bar (second) is a Buy, the proceeds will be available to make the purchase?

Overall, for very low price changes on short time scales (seconds, ticks) you have trade thousands of shares for each alert in order to be profitable. But trading so many shares on a short time scale can result in a lack of sufficient funds very quickly since sell and cover trades won't get filled as quickly as the unit of time scale you are on.

Next, if you are training a deep learning ANN with several hidden layers, which will require hundreds of iterations to determine whether to go long (Buy) or short for each second's bar, how are you going to be able to run your input through the ANNs architecture and trained connection weights to make the prediction before the next bar?

Overall, there are some trading issues which could be disastrous based on what you want to do, such that trade settlements will become clogged up and insufficient funds will occur for future trades (selling shares can take many seconds).

• I don't want to trade at very low price changes. I would like to trade with the machine as a normal trader would. My market making strategy is order flow, littlefishfx.com/forex-education/…. How with python could I tell which price would be sell, buy or do nothing (1, -1, 0)? Apr 14, 2018 at 19:00
• @Jeremie Your comment makes it sound more like you want a set of rules that mimic what a professional trader thinks, rather than some kind of machine learning prediction.
– Dave
Jul 14, 2021 at 18:09